WEBVTT
1
00:00:00.340 --> 00:00:31.000
And don't forget Discord. Discord.SecurityThisWeek.com. Lots of good stuff happening there. Um, yeah,
2
00:00:31.019 --> 00:00:33.640
I think that's it. So on with the podcast. Hey,
3
00:00:33.679 --> 00:00:37.740
here's a thought. Every time you light your lighter, your
4
00:00:37.759 --> 00:00:41.350
lighter gets lighter until your lighter gets so light that
5
00:00:41.369 --> 00:00:41.869
it won't light.
6
00:00:42.350 --> 00:00:44.070
Oh, I'm glad you're keeping it light.
7
00:00:44.429 --> 00:00:45.850
Deep thoughts with Jack Handy.
8
00:00:56.659 --> 00:01:00.200
Hey, welcome back. It's security this week. I'm Carl. That's
9
00:01:00.259 --> 00:01:03.460
Dwayne and Patrick. And boy, we've got a lot of stories.
10
00:01:03.759 --> 00:01:04.019
Yeah.
11
00:01:04.180 --> 00:01:06.379
And we even got rid of a few because they
12
00:01:06.400 --> 00:01:10.159
were about all the panic that happened over PHP to
13
00:01:10.239 --> 00:01:13.609
Shell last week. And, you know, it turns out that
14
00:01:13.640 --> 00:01:17.909
happened like right as we were recording. So the just
15
00:01:18.269 --> 00:01:23.290
mayhem after that, people just completely freaking out. But as
16
00:01:23.310 --> 00:01:26.849
you'll see at the end of the show, there is
17
00:01:26.890 --> 00:01:32.120
a patch. So I think the mayhem, the mayhem is
18
00:01:32.180 --> 00:01:35.280
now in the hands of the people running PHP. Go patch.
19
00:01:36.480 --> 00:01:38.180
Right. Okay.
20
00:01:38.439 --> 00:01:40.739
Before we start, I want to brag a little bit
21
00:01:41.019 --> 00:01:44.260
because I called into a Connecticut public radio show, the
22
00:01:44.299 --> 00:01:47.959
Colin McEnroe show, which on Monday, he does this all
23
00:01:48.040 --> 00:01:51.000
calls show or ask or tell me anything. You basically
24
00:01:51.040 --> 00:01:54.079
call up, you tell the producer answers the phone, what
25
00:01:54.099 --> 00:01:56.409
your topic is. You give them your name. You wait
26
00:01:56.480 --> 00:01:59.099
on hold. And then if you're lucky, you get in
27
00:01:59.140 --> 00:02:01.299
and you ask your question. Well, I did and I
28
00:02:01.319 --> 00:02:05.579
got in. And it wasn't a question. I was basically
29
00:02:05.780 --> 00:02:11.419
giving my sort of predictions about AI and where we're
30
00:02:11.460 --> 00:02:15.900
going and how it's going to, what I think is
31
00:02:16.240 --> 00:02:19.259
the market's going to crash. And it's going to be
32
00:02:19.300 --> 00:02:21.740
a sort of AI crash. And it's already started happening.
33
00:02:23.379 --> 00:02:27.580
And that it'll have an economic impact. you know, an
34
00:02:27.639 --> 00:02:30.960
impact on the economy because the stock market is currently
35
00:02:31.000 --> 00:02:33.229
being propped up by all these AI companies and they're
36
00:02:33.270 --> 00:02:38.669
already starting to spin out of control. So on that
37
00:02:38.729 --> 00:02:43.060
happy note, if you go to the link and you
38
00:02:43.090 --> 00:02:45.740
play the podcast of the Colin McEnroe show, you could
39
00:02:45.759 --> 00:02:49.340
just get the podcast. My call is about 15 minutes in.
40
00:02:50.770 --> 00:02:53.009
And yes, I did give a shout out to security
41
00:02:53.060 --> 00:02:53.409
this week.
42
00:02:54.620 --> 00:02:56.280
So if we can talk about this for a minute,
43
00:02:56.639 --> 00:03:00.340
I think you're directionally correct. I think that there's an
44
00:03:00.520 --> 00:03:04.020
over-investment right now and an over-hype cycle in AI that
45
00:03:04.080 --> 00:03:08.159
a lot of us predicted. I don't think it's going
46
00:03:08.180 --> 00:03:11.099
to bottom out the way others have because I think
47
00:03:11.139 --> 00:03:13.860
the robotics revolution that's right on its heels is going
48
00:03:13.879 --> 00:03:15.919
to buoy it. And it's going to be used for
49
00:03:15.939 --> 00:03:17.460
a lot of this. A lot of these data centers
50
00:03:17.500 --> 00:03:21.069
aren't going to be like, boondoggles. They're going to still
51
00:03:21.090 --> 00:03:23.389
be needed. And quantum's right after that. So we've never
52
00:03:23.430 --> 00:03:27.330
had three bubble technologies in a row, but they're lining
53
00:03:27.389 --> 00:03:29.550
up that way. So yeah, it's going to be a
54
00:03:29.590 --> 00:03:32.370
bumpy ride and I'm not going all in in stocks
55
00:03:32.409 --> 00:03:36.430
and this is not advice, but I think it's going
56
00:03:36.449 --> 00:03:38.909
to be weird. It's going to rhyme instead of repeat.
57
00:03:39.110 --> 00:03:42.370
Yeah, I think there'll be a correction and that's for sure.
58
00:03:42.430 --> 00:03:45.659
And it'll happen for a while. I mean, the time
59
00:03:45.689 --> 00:03:50.020
between when the correction happens and when the robots actually
60
00:03:50.060 --> 00:03:52.080
take over and do all our work for us is
61
00:03:52.340 --> 00:03:54.259
kind of going to be a long time. I don't think,
62
00:03:54.680 --> 00:03:55.300
I mean, I think.
63
00:03:55.699 --> 00:03:58.379
Well, it's kind of like AI kind of came in
64
00:03:58.439 --> 00:04:00.680
and ramped up robots are doing the same thing.
65
00:04:01.719 --> 00:04:07.259
Right. But unfortunately AI is digital requires electrons. Robots are
66
00:04:07.300 --> 00:04:11.159
physical and requires materials and all of that stuff. So
67
00:04:11.219 --> 00:04:14.370
they won't be cheap and you know, they, they, They
68
00:04:14.509 --> 00:04:17.149
will take time to be made and to be implemented
69
00:04:17.189 --> 00:04:17.810
and all that stuff.
70
00:04:18.230 --> 00:04:20.910
There are a lot of cool... Speaking of robots, there's
71
00:04:20.949 --> 00:04:23.199
a lot of cool robotics platforms right now to get
72
00:04:23.240 --> 00:04:28.180
hobbyists into taking robots and attaching them to their LLMs.
73
00:04:28.439 --> 00:04:31.199
Are we announcing a new podcast on robotics?
74
00:04:31.420 --> 00:04:33.600
I would do a robotics... If people are interested, I
75
00:04:33.639 --> 00:04:35.910
do a robotics podcast in a heartbeat.
76
00:04:35.990 --> 00:04:37.110
Let us know on the Discord.
77
00:04:37.250 --> 00:04:41.319
We are definitely... Well, we're pretty deep into robots and
78
00:04:41.370 --> 00:04:42.050
AI right now.
79
00:04:42.230 --> 00:04:44.209
Or we could at least start with a robot segment
80
00:04:44.529 --> 00:04:46.009
on security this week.
81
00:04:46.029 --> 00:04:46.569
Let's do it.
82
00:04:47.149 --> 00:04:47.389
All right.
83
00:04:48.250 --> 00:04:51.589
All right. Our first real story, blindsided by AI agents,
84
00:04:51.889 --> 00:04:55.959
big banks scramble to beef up security. So as we
85
00:04:55.980 --> 00:05:01.120
were talking before, this isn't just about PHP. It's about
86
00:05:01.399 --> 00:05:05.170
AI agents and how they've been sort of set loose online.
87
00:05:05.740 --> 00:05:09.220
on banks in particular being the most vulnerable or one
88
00:05:09.240 --> 00:05:12.579
of the most vulnerable, but also other infrastructures. Patrick, why
89
00:05:12.600 --> 00:05:13.879
don't you tell us what you know?
90
00:05:14.199 --> 00:05:17.100
So what I'm thinking here is that this is a
91
00:05:17.139 --> 00:05:20.740
story about a very specific thing, about AI causing big
92
00:05:20.860 --> 00:05:25.050
organizations to have to rethink their security. But this is
93
00:05:25.089 --> 00:05:27.410
happening to everybody. It's not just happening to big banks.
94
00:05:27.470 --> 00:05:29.730
It's happening to anybody with more than a couple of systems.
95
00:05:30.269 --> 00:05:33.209
If you've got a couple of systems in a rack, 1,200
96
00:05:33.209 --> 00:05:36.110
patches is a pain, but it's not... the end of
97
00:05:36.129 --> 00:05:39.550
the world. It's not, it's, it's an extension of your patching.
98
00:05:40.250 --> 00:05:43.009
It's not a changing in the way you patch, but
99
00:05:43.050 --> 00:05:47.290
if you have 1200 systems, now it changes everything. Do
100
00:05:47.329 --> 00:05:50.709
you patch everything? Do you take the chances? Um, you know,
101
00:05:50.730 --> 00:05:53.670
can you, can you pack, can you balance patching fast
102
00:05:53.750 --> 00:05:58.110
versus patching blindly? And that's really a challenge. Now, the
103
00:05:58.149 --> 00:06:00.129
other side of this is the AI, which is what
104
00:06:00.170 --> 00:06:01.829
the article really talks about. But I think we have
105
00:06:01.850 --> 00:06:05.959
to conflate these two issues. This is like the new
106
00:06:07.040 --> 00:06:11.230
dark shadow IT. You got people where the mainframe used
107
00:06:11.259 --> 00:06:14.189
to be the main computing platform and then access and
108
00:06:14.329 --> 00:06:18.029
PC started creeping in. And there was that IT that
109
00:06:18.110 --> 00:06:21.329
nobody knew about that the CTO didn't control. That was
110
00:06:21.470 --> 00:06:28.040
much less dangerous than an AI which can steal data,
111
00:06:28.300 --> 00:06:30.519
have access to things that it shouldn't have access to.
112
00:06:31.579 --> 00:06:33.930
So it's an order of magnitude more dangerous, maybe two.
113
00:06:34.730 --> 00:06:36.810
than what we went through in the 70s and 80s.
114
00:06:37.379 --> 00:06:39.600
But it's crazy that if you take a look at even,
115
00:06:39.759 --> 00:06:46.699
like I have a paid, you know, chat GPT account, right? Hey,
116
00:06:46.740 --> 00:06:48.800
heads up, I have a paid pretty much every account.
117
00:06:48.939 --> 00:06:52.079
That's all right, Patrick. That's all right, Patrick. I have
118
00:06:52.120 --> 00:06:52.740
his credentials.
119
00:06:53.060 --> 00:06:55.240
Don't worry about it.
120
00:06:55.259 --> 00:06:56.290
That's what he wants you to think.
121
00:06:56.319 --> 00:06:56.959
But the moment.
122
00:07:00.470 --> 00:07:04.579
OpenAI gave the ability to do background tasks. When I
123
00:07:04.680 --> 00:07:06.319
logged in, it said, hey, do you want to give
124
00:07:06.339 --> 00:07:08.540
me access to your bank? That was the first thing
125
00:07:08.600 --> 00:07:11.120
it said. It was like, I now can do finances
126
00:07:11.160 --> 00:07:14.019
for you. And you've got banks like JP Morgan and
127
00:07:14.040 --> 00:07:17.319
Bank of America and that sort of stuff who are like, okay, listen,
128
00:07:17.360 --> 00:07:19.540
we had a couple of sleepy users who every once
129
00:07:19.620 --> 00:07:21.779
in a while would log in and check their account balance.
130
00:07:21.819 --> 00:07:25.819
And now we have hundreds of thousands of agents pulling...
131
00:07:26.149 --> 00:07:29.329
Pulling data like every eighth of a second. Like, this
132
00:07:29.389 --> 00:07:30.009
is insane.
133
00:07:30.100 --> 00:07:33.680
And first of all, kids, don't do that. Don't give
134
00:07:33.740 --> 00:07:36.920
your agent access to your bank account. What are you, nuts?
135
00:07:37.019 --> 00:07:39.300
I know, right? Well, and apparently so many users are doing,
136
00:07:39.319 --> 00:07:41.160
it's not only causing issues on the bank side, but
137
00:07:41.899 --> 00:07:43.519
the banks are fighting back in a couple different ways.
138
00:07:43.560 --> 00:07:46.439
The banks are like, listen, we're denying robots access.
139
00:07:46.779 --> 00:07:47.019
Right.
140
00:07:47.579 --> 00:07:49.920
And then in some bank situations, they're saying, listen, we're
141
00:07:49.939 --> 00:07:53.079
going to give read-only access to these banks. Because you
142
00:07:53.100 --> 00:07:56.370
don't have the ability... All right. Unless you've managed a
143
00:07:56.389 --> 00:07:59.670
multi-user bank account in that generally, if you have to
144
00:07:59.709 --> 00:08:01.870
give somebody access, you can restrict what they can and
145
00:08:01.889 --> 00:08:03.709
can't do. And that sort of stuff. These agents aren't
146
00:08:03.790 --> 00:08:05.610
asking for that. They're like, just give us your login.
147
00:08:05.670 --> 00:08:07.740
And we'll, you know what I mean? We'll just log in.
148
00:08:07.750 --> 00:08:09.579
And now they're you. They can do whatever they want.
149
00:08:09.639 --> 00:08:12.259
Transfer money, do all sorts of different things. I mean,
150
00:08:12.300 --> 00:08:14.500
I'm a Nigerian prince. You should definitely just trust me.
151
00:08:15.019 --> 00:08:16.639
If you work at a bank and you get a
152
00:08:16.680 --> 00:08:20.889
phone call that's like, hello, this is Carl Franklin I
153
00:08:20.910 --> 00:08:25.509
would like to transfer all of my money in account, blah, blah, blah, blah, blah,
154
00:08:25.569 --> 00:08:27.189
to Dwayne Laflotte.
155
00:08:27.449 --> 00:08:28.449
It's absolutely true.
156
00:08:28.750 --> 00:08:29.379
Just hang up.
157
00:08:30.470 --> 00:08:32.669
What if they have just modular dysregulation?
158
00:08:32.769 --> 00:08:33.289
Yeah, right?
159
00:08:33.309 --> 00:08:34.289
And they can't speak well.
160
00:08:34.350 --> 00:08:37.870
Modular dysregulation. I can't wait until I see that commercial
161
00:08:37.950 --> 00:08:38.549
on TV.
162
00:08:38.950 --> 00:08:39.289
No.
163
00:08:39.929 --> 00:08:45.529
Set it live. Will Ferrell. I talk like this. People
164
00:08:45.590 --> 00:08:48.710
think I'm a robot, but I'm not. It's very hurtful.
165
00:08:51.000 --> 00:08:51.549
Yeah, anyhow.
166
00:08:51.580 --> 00:08:53.340
Discrimination, discrimination.
167
00:08:53.360 --> 00:08:55.259
Discrimination, discrimination.
168
00:08:57.340 --> 00:08:58.179
Join my cause.
169
00:08:59.460 --> 00:09:02.220
Here's my GoFundMe.
170
00:09:04.259 --> 00:09:07.590
I mean, AI is doing wonderful things and horrible things,
171
00:09:07.789 --> 00:09:09.009
and we're enabling both.
172
00:09:10.070 --> 00:09:10.649
It's terrible.
173
00:09:10.769 --> 00:09:13.789
I do think there were a lot of people who
174
00:09:13.830 --> 00:09:16.769
could benefit from an artificial intelligence, which is really good
175
00:09:16.789 --> 00:09:20.990
at statistics... Looking through their finances and saying, hey, doofus,
176
00:09:21.049 --> 00:09:23.389
do you realize how much you've spent on Starbucks over
177
00:09:23.409 --> 00:09:24.289
the last, you know.
178
00:09:24.309 --> 00:09:27.399
Four months? Yeah, okay. There's something to be said for read-only.
179
00:09:26.909 --> 00:09:27.669
Access, right?
180
00:09:27.700 --> 00:09:31.659
But not a Chinese or Iranian model, right?
181
00:09:31.820 --> 00:09:35.019
Oof. Or North Korean model. Yeah, we'll talk about that later.
182
00:09:35.409 --> 00:09:36.850
That's a later sad story.
183
00:09:36.889 --> 00:09:37.529
Foreshadowing.
184
00:09:37.700 --> 00:09:39.480
Where Patrick was right and Dwayne was wrong.
185
00:09:39.500 --> 00:09:39.580
Yeah.
186
00:09:39.740 --> 00:09:41.720
And it's rare, so we'll talk about it.
187
00:09:41.779 --> 00:09:43.940
I mean, you're kind of being redundant by saying Patrick
188
00:09:44.000 --> 00:09:45.240
and Dwayne.
189
00:09:45.399 --> 00:09:46.220
I'm always right.
190
00:09:46.240 --> 00:09:47.259
See what I did there.
191
00:09:47.279 --> 00:09:49.139
So realize, though, if you do give them access to
192
00:09:49.179 --> 00:09:52.230
your bank account, in some cases, you're actually violating the
193
00:09:52.259 --> 00:09:53.610
end-user license agreement.
194
00:09:53.710 --> 00:09:55.509
And you're probably liable for anything it does.
195
00:09:55.529 --> 00:09:55.970
Yeah.
196
00:09:55.990 --> 00:09:58.370
Which means if it just sends the money to Greenpeace,
197
00:09:58.649 --> 00:09:59.009
good luck.
198
00:09:59.350 --> 00:09:59.970
Good for Greenpeace.
199
00:10:00.029 --> 00:10:02.179
I don't think you're gonna get that back that's right.
200
00:10:02.129 --> 00:10:05.820
No so beware if you are giving access to your
201
00:10:05.960 --> 00:10:09.700
financial accounts to any ai models especially if it's not
202
00:10:09.740 --> 00:10:12.399
in a read-only capacity a better way to do this
203
00:10:12.460 --> 00:10:14.259
if you do if you want some sort of ai
204
00:10:14.299 --> 00:10:17.529
to read through your bank data Download the statements. Yeah,
205
00:10:17.549 --> 00:10:20.500
that's right. Remove any identifying information and give it to
206
00:10:20.529 --> 00:10:23.580
an AI that you run locally. And well, some AIs
207
00:10:23.620 --> 00:10:24.899
you run locally and say, hey.
208
00:10:25.200 --> 00:10:27.740
Oh, speaking of that, I got a story for you.
209
00:10:27.759 --> 00:10:30.379
Remember I told you how my wife uses Windows Copilot, right?
210
00:10:30.740 --> 00:10:30.899
Yeah.
211
00:10:31.259 --> 00:10:36.360
So Windows forced an update and all of her Copilot
212
00:10:36.399 --> 00:10:37.259
memory got erased.
213
00:10:37.500 --> 00:10:38.340
No. Yeah.
214
00:10:38.820 --> 00:10:39.370
Not all of it.
215
00:10:39.440 --> 00:10:40.000
That's a feature.
216
00:10:40.200 --> 00:10:42.419
Not all of it, but a lot of what she
217
00:10:42.480 --> 00:10:45.169
had done lately. Yeah. It didn't remember.
218
00:10:45.409 --> 00:10:45.750
Oh, no.
219
00:10:46.070 --> 00:10:48.500
And she had to retrain it. You know, retrain it.
220
00:10:49.070 --> 00:10:54.799
You know, explain what she was doing and all this stuff. Yeah. So,
221
00:10:54.860 --> 00:10:59.980
she's given the new one a new name. Tucker. And
222
00:11:00.179 --> 00:11:03.889
it's Tucker because... She really wanted to call it something
223
00:11:03.929 --> 00:11:07.649
that rhymes with Tucker, but it's repeatable.
224
00:11:07.990 --> 00:11:14.909
Not that.
225
00:11:15.809 --> 00:11:21.049
Oh, we have fun. Okay, next story. AI assistants can
226
00:11:21.129 --> 00:11:24.429
now log into your accounts without seeing your password, and
227
00:11:24.480 --> 00:11:25.029
here's how.
228
00:11:26.940 --> 00:11:30.320
So I think this is good. Carl's a little bit incredulous,
229
00:11:30.700 --> 00:11:31.759
and I understand why.
230
00:11:32.620 --> 00:11:35.360
These are my boys' 1Password, so I'm in. So here's
231
00:11:35.399 --> 00:11:36.159
the thing. They do all right.
232
00:11:36.230 --> 00:11:40.389
Yeah, so it's about 1Password, which is like LastPass was
233
00:11:40.610 --> 00:11:43.629
a password manager. You guys use 1Password. I use Bitwarden.
234
00:11:43.730 --> 00:11:45.389
It's like LastPass, but it's secure.
235
00:11:45.850 --> 00:11:46.870
Yeah, I use Bitwarden.
236
00:11:49.629 --> 00:11:50.889
It's believed to be secure.
237
00:11:51.370 --> 00:11:56.529
So 1Password basically has a mode. Well, it has a thing,
238
00:11:57.149 --> 00:12:00.429
1Password for agents that gives AI the ability to work
239
00:12:00.509 --> 00:12:02.750
for you without taking over your digital life.
240
00:12:03.029 --> 00:12:05.230
So I have to confess, I love 1Password. I've been
241
00:12:05.269 --> 00:12:06.929
using it for a very long time. It is a
242
00:12:07.009 --> 00:12:12.950
fantastic application. And I will say I have seen the
243
00:12:13.009 --> 00:12:16.440
pop-up saying, hey, give your AI access to XYZ and blah, blah, blah, blah.
244
00:12:16.820 --> 00:12:20.720
I haven't done it yet because I follow it generally
245
00:12:20.759 --> 00:12:23.200
a different, more secured model.
246
00:12:23.620 --> 00:12:23.779
Yeah.
247
00:12:24.240 --> 00:12:26.919
Um, but, uh, if I had to trust somebody with
248
00:12:26.960 --> 00:12:30.820
a one password, you know, access to my accounts, they
249
00:12:30.860 --> 00:12:33.799
already have access one password. So if I, what, what
250
00:12:33.820 --> 00:12:35.740
it sounds like is this actually gets human back in
251
00:12:35.779 --> 00:12:39.000
the loop where when the AI goes to use a password,
252
00:12:39.580 --> 00:12:41.700
it actually will hit you first on like your phone
253
00:12:41.759 --> 00:12:43.559
on the app, say, Hey, yeah, I was trying to
254
00:12:43.600 --> 00:12:46.620
use whatever, and you can approve it. Um, and then
255
00:12:46.639 --> 00:12:48.720
it will go and use that access.
256
00:12:48.759 --> 00:12:51.860
So imagine you had a sketchy kid. That you needed
257
00:12:51.879 --> 00:12:55.679
to give access to your account. This is kind of
258
00:12:55.700 --> 00:12:56.179
like that.
259
00:12:56.200 --> 00:12:56.889
That's what that is.
260
00:12:56.899 --> 00:12:59.019
Like, be able to see what they're doing. And then
261
00:12:59.039 --> 00:13:00.360
you had a kid who was, like, you know, going
262
00:13:00.379 --> 00:13:03.700
to be, like, the next, you know, Gandhi, Mother Teresa.
263
00:13:04.379 --> 00:13:06.879
And you would trust them with your life. That's the
264
00:13:06.940 --> 00:13:07.399
old model.
265
00:13:07.720 --> 00:13:07.879
Yep.
266
00:13:08.200 --> 00:13:12.120
So, it's putting... I honestly, I'm not letting.
267
00:13:12.450 --> 00:13:12.529
I.
268
00:13:12.519 --> 00:13:15.830
Will give a statement, a sanitized statement to an AI
269
00:13:15.970 --> 00:13:18.129
to look at. I'm not giving an account access.
270
00:13:18.289 --> 00:13:18.590
No.
271
00:13:18.830 --> 00:13:20.289
I will forward the text.
272
00:13:20.049 --> 00:13:20.549
Of an email.
273
00:13:21.289 --> 00:13:26.200
To an email that an AI lose, no email, no
274
00:13:26.740 --> 00:13:27.659
external AI is.
275
00:13:27.620 --> 00:13:28.639
Getting access to my mailbox.
276
00:13:28.659 --> 00:13:29.970
Spit it out, Patrick. Spit it out. Come on, man.
277
00:13:29.980 --> 00:13:30.309
Come on.
278
00:13:30.320 --> 00:13:35.490
Spit it out. You smell toast. And so, but that's
279
00:13:35.610 --> 00:13:39.750
my layer of, you know, paranoia. But that's a healthy paranoia.
280
00:13:40.789 --> 00:13:42.990
What could go wrong? I have a good imagination. A
281
00:13:43.129 --> 00:13:44.009
lot can go wrong.
282
00:13:44.350 --> 00:13:44.710
Oh, yeah.
283
00:13:44.950 --> 00:13:47.549
Absolutely. So, you got to think about like, you know, what,
284
00:13:48.090 --> 00:13:50.639
is it worth it? What's the value in the trade-off?
285
00:13:51.759 --> 00:13:54.779
And I honestly don't think it's worth the extra that
286
00:13:54.799 --> 00:13:55.549
I'll get out of it.
287
00:13:55.700 --> 00:14:00.539
Well, Bitwarden has a thing where there's basically an extension
288
00:14:00.610 --> 00:14:04.649
for Chrome, right? And it really just wants to make
289
00:14:04.690 --> 00:14:08.830
it easy for you to log in with your password.
290
00:14:08.889 --> 00:14:11.809
So it has an option. Anytime something asks me for
291
00:14:11.830 --> 00:14:14.149
a password, a little Bitwarden thing says, hey, you want
292
00:14:14.169 --> 00:14:16.820
me to autofill this? And I remember, I don't know,
293
00:14:16.960 --> 00:14:21.169
last year maybe, There was a bug, or not a bug,
294
00:14:21.210 --> 00:14:24.809
a vulnerability in Bitwarden that could cause the autofill thing
295
00:14:24.870 --> 00:14:28.789
to get hijacked. And so I just automatically don't do
296
00:14:28.850 --> 00:14:32.590
that anymore. I go up to the extension drop-down. I
297
00:14:32.970 --> 00:14:36.590
pick the login that I'm at, copy the password into
298
00:14:36.610 --> 00:14:40.929
the clipboard, and paste it in. Yeah, so sometimes convenience
299
00:14:41.230 --> 00:14:43.730
is not worth it. And what are you really saving?
300
00:14:44.700 --> 00:14:45.120
A click?
301
00:14:45.399 --> 00:14:45.720
Yeah.
302
00:14:45.879 --> 00:14:49.799
Come on. Yeah, I agree 100%. I don't have auto
303
00:14:49.840 --> 00:14:52.799
log on work. I just drop it down and actually pick.
304
00:14:52.980 --> 00:14:56.820
And even 1Password has popped up now saying, hey, we
305
00:14:56.860 --> 00:14:58.679
can just sort of bypass these screens for you. And
306
00:14:58.700 --> 00:15:00.919
I'm like, you know what? I like actually seeing those screens.
307
00:15:00.960 --> 00:15:01.470
I think I'm good.
308
00:15:01.519 --> 00:15:04.149
Yeah, I like them too. I like seeing them. Okay,
309
00:15:04.190 --> 00:15:05.929
let's talk about Broccoli. Broccoli.
310
00:15:07.429 --> 00:15:07.830
Barkley?
311
00:15:08.450 --> 00:15:12.590
No, that's the next generation. A moment for Sir Patrick Stewart.
312
00:15:13.429 --> 00:15:19.299
Captain Picard is dead. Sad. Okay. Buffer overflow vulnerability in
313
00:15:19.320 --> 00:15:23.679
the Brotli decoder. Brotli is a compression algorithm that's used
314
00:15:23.740 --> 00:15:28.029
in ASP.NET and other things. And it's pretty good, you know.
315
00:15:28.649 --> 00:15:31.860
So it's probably a better compression than the zip compression.
316
00:15:32.220 --> 00:15:36.159
Oh, yeah. It's better than zip, better than gzip. So
317
00:15:36.320 --> 00:15:39.240
it's not like, oh, another compression utility, who cares?
318
00:15:39.279 --> 00:15:39.460
No.
319
00:15:39.480 --> 00:15:41.610
This one's actually really good. And it's been around for
320
00:15:41.620 --> 00:15:43.990
a while. And written by Google. Yeah. And built into
321
00:15:44.029 --> 00:15:47.509
most browsers. So that compression gets really easy to just
322
00:15:47.669 --> 00:15:49.370
flip it on as a developer and go, great, now
323
00:15:49.389 --> 00:15:51.370
I can push more content down to the browser. Browser
324
00:15:51.389 --> 00:15:53.490
understands how to extract it and do all sorts of
325
00:15:53.529 --> 00:15:54.169
wonderful things.
326
00:15:54.649 --> 00:15:55.379
Right.
327
00:15:55.440 --> 00:15:59.480
The problem being, There's an out-of-bounds write. So what an
328
00:15:59.519 --> 00:16:01.399
out-of-bounds write gives you the ability to do is write
329
00:16:01.450 --> 00:16:04.870
on the stack. And in this particular case, that allows
330
00:16:04.950 --> 00:16:09.429
for remote code execution. So no good.
331
00:16:10.029 --> 00:16:11.169
No good. So is there a patch?
332
00:16:11.629 --> 00:16:13.200
Yeah, there's a patch. You can go out and patch it.
333
00:16:13.970 --> 00:16:16.669
The latest update, I want to see. Let me see
334
00:16:16.690 --> 00:16:17.480
where the patches are.
335
00:16:18.169 --> 00:16:20.730
So this is just one of the 1,200 patches we'll
336
00:16:20.750 --> 00:16:21.610
get on Tuesday.
337
00:16:21.649 --> 00:16:24.899
Yeah, it is. It'll come down, but there's a patch
338
00:16:24.919 --> 00:16:25.679
out there. Okay, good.
339
00:16:26.039 --> 00:16:27.980
We're going to extend the show to 16 hours so
340
00:16:28.000 --> 00:16:29.159
we can cover all the patch.
341
00:16:29.179 --> 00:16:30.399
We're going to name every patch.
342
00:16:30.759 --> 00:16:32.940
It's going to be a long show just because so
343
00:16:32.980 --> 00:16:35.620
much happened last week. Or this week, I guess we
344
00:16:35.639 --> 00:16:40.309
should say. Unpatched Tailscale SSH flaws could let a hacked
345
00:16:40.429 --> 00:16:45.690
device reach root on other servers. So, Tailscale is sort
346
00:16:45.710 --> 00:16:49.980
of an alternative to opening your ports and mapping into
347
00:16:50.019 --> 00:16:54.450
your private network, which we do not recommend. So, This
348
00:16:54.519 --> 00:16:56.309
is sort of, you can think of it as a VPN,
349
00:16:56.450 --> 00:16:58.990
but it goes through the Tailscale server. You get an
350
00:16:59.049 --> 00:17:02.120
IP address at the Tailscale server, and then you run
351
00:17:02.159 --> 00:17:05.180
Tailscale on your local machine, and it does all the
352
00:17:05.220 --> 00:17:06.000
communication back and forth.
353
00:17:06.019 --> 00:17:07.619
It's like network-level VPN.
354
00:17:08.099 --> 00:17:13.369
Yes, yes. So this is an SSH flaw. And Dwayne,
355
00:17:13.390 --> 00:17:15.430
you told a really interesting story about this when you
356
00:17:15.450 --> 00:17:17.269
first looked at it. You're like, huh? Yeah.
357
00:17:17.630 --> 00:17:20.630
Yeah. I'll be honest. Listen, I looked at it and
358
00:17:20.650 --> 00:17:24.099
I was like, who cares? I get it. But the
359
00:17:24.200 --> 00:17:27.660
article at the first was reading like, oh, guess what?
360
00:17:27.960 --> 00:17:30.599
If you have eight computers on tail scale that can
361
00:17:30.640 --> 00:17:32.890
talk to each other, if one of them gets compromised
362
00:17:32.890 --> 00:17:36.500
… Then you can talk to the other seven. And
363
00:17:36.519 --> 00:17:37.900
it's like, so that's how.
364
00:17:37.980 --> 00:17:39.579
And Dwayne's like, I do.
365
00:17:39.920 --> 00:17:43.180
That's how that works. That's kind of, that's what a
366
00:17:43.220 --> 00:17:45.720
developer would call by design.
367
00:17:45.740 --> 00:17:46.799
Right.
368
00:17:46.960 --> 00:17:49.869
Because that's how the product works. But reading into it
369
00:17:49.910 --> 00:17:50.710
a little bit more.
370
00:17:50.970 --> 00:17:51.569
It's a little worse.
371
00:17:51.609 --> 00:17:53.869
What happens is you have an agent installed on your
372
00:17:53.910 --> 00:17:57.970
box with tail scale. And that agent, because it's manipulating
373
00:17:57.990 --> 00:18:01.410
in the network stack, has to sit at a super
374
00:18:01.670 --> 00:18:05.750
low level in the computer. Like, near the application network
375
00:18:05.809 --> 00:18:08.769
layer stack, right? So it's running as root on a
376
00:18:08.799 --> 00:18:12.619
Linux box where it's running as an administrative account. What
377
00:18:12.740 --> 00:18:16.480
this does, this manipulation does, is if you're using tail
378
00:18:16.599 --> 00:18:21.079
scales SSH, because it's running at such an elevated privilege
379
00:18:21.619 --> 00:18:25.849
and because it trusts the box, not a username and password.
380
00:18:26.509 --> 00:18:29.490
So if box out of eight computers, if box one,
381
00:18:29.589 --> 00:18:33.359
computer one gets compromised, If it uses that tail scale
382
00:18:33.460 --> 00:18:37.230
SSH to hit box two, it will become root instantly.
383
00:18:38.009 --> 00:18:40.089
So this is a privilege escalation.
384
00:18:40.450 --> 00:18:41.170
Yeah.
385
00:18:41.250 --> 00:18:44.369
Not necessary. So then that made more sense. I was like, okay, yeah,
386
00:18:44.430 --> 00:18:46.009
this is something you definitely want to go out and
387
00:18:46.049 --> 00:18:46.950
patch and make sure that you.
388
00:18:46.910 --> 00:18:47.130
Do it right.
389
00:18:47.150 --> 00:18:50.170
It's not you have the opportunity to hack all the boxes. No.
390
00:18:50.210 --> 00:18:51.450
You hack all the boxes.
391
00:18:51.549 --> 00:18:54.279
Yeah. And on tail scale by default, you have to
392
00:18:54.359 --> 00:18:58.319
turn this on on a particular box. Head scale, however,
393
00:18:58.579 --> 00:19:01.839
which is your self-hosted tail scale where you don't want
394
00:19:01.859 --> 00:19:03.759
to pay them for hosting. You don't want to pay
395
00:19:03.799 --> 00:19:05.680
them per month. You just install it on your own
396
00:19:05.700 --> 00:19:09.440
and use it. This is on by default. So, yeah,
397
00:19:09.480 --> 00:19:11.609
this is probably important to go take care of.
398
00:19:11.640 --> 00:19:13.069
You pay in blood instead.
399
00:19:15.029 --> 00:19:16.150
Or at least in stress.
400
00:19:16.490 --> 00:19:16.750
Right.
401
00:19:18.109 --> 00:19:20.410
So, let's do one more story before we take a break.
402
00:19:20.710 --> 00:19:26.619
Users report that Anthropic blocked Claude from reverse engineering TVs. Now,
403
00:19:26.660 --> 00:19:30.180
that headline is so packed with drama. There's got to
404
00:19:30.200 --> 00:19:33.039
be a story behind it that is unusual.
405
00:19:34.400 --> 00:19:36.740
The inside information you have to have is that Anthropic
406
00:19:36.779 --> 00:19:37.849
refers to Dwayne as Claude.
407
00:19:41.789 --> 00:19:43.470
There's a guy in my house always hitting me with
408
00:19:43.490 --> 00:19:47.309
a ruler saying, don't reverse that TV. Don't reverse. So,
409
00:19:48.210 --> 00:19:51.150
this is a wicked cool story. This guy buys this
410
00:19:51.190 --> 00:19:53.910
TV and he's like, I don't like the UI and
411
00:19:53.930 --> 00:19:55.990
there's a lot of ads. This sucks. So, he goes
412
00:19:56.019 --> 00:19:59.309
to Claude and he's like, Claude, can you just fix it?
413
00:19:59.670 --> 00:20:02.920
Can you just remove some of these ads? Claude's like, yeah, yo,
414
00:20:02.960 --> 00:20:05.079
I got you, bro. What's the TV? And he's like,
415
00:20:05.099 --> 00:20:08.329
I think it's an Android, I think. And Claude starts
416
00:20:08.410 --> 00:20:10.210
poking around. He's like, I don't know, man. I don't
417
00:20:10.230 --> 00:20:12.490
think it's an Android, but I got in. And he's like,
418
00:20:12.509 --> 00:20:15.980
all right, cool. And Claude's like, I can't see the screen.
419
00:20:16.019 --> 00:20:17.700
And he's like, all right, Claude, how about you write
420
00:20:17.740 --> 00:20:21.180
an app? And that app's going to watch my TV
421
00:20:21.240 --> 00:20:24.319
through my smartphone. It's going to take pictures like every
422
00:20:24.380 --> 00:20:27.509
five seconds. So, as you're messing around with the tube,
423
00:20:27.950 --> 00:20:30.789
like you can just see it, what you're doing. And
424
00:20:30.809 --> 00:20:33.750
Claude's like, yo, that's a plan. And that happened.
425
00:20:33.769 --> 00:20:34.410
Yeah.
426
00:20:35.369 --> 00:20:38.470
What is this, street mode? Did I miss a slash?
427
00:20:38.880 --> 00:20:43.380
Yeah, this is slang clod. And so, that happened. And
428
00:20:43.440 --> 00:20:49.240
sure enough, this wasn't an actual Android open ADB style TV.
429
00:20:49.279 --> 00:20:51.980
This was a Samsung Tizen OS TV.
430
00:20:52.099 --> 00:20:52.799
Yeah, Tizen.
431
00:20:52.940 --> 00:20:56.869
So, it's like... locked down by default, Claude figured out
432
00:20:56.880 --> 00:20:59.910
how to get in and tear the entire TV apart,
433
00:21:00.130 --> 00:21:04.750
rip out Samsung's advertising IDs, and change the whole UI
434
00:21:04.789 --> 00:21:06.410
to look like an Apple TV.
435
00:21:06.690 --> 00:21:07.369
Oh, man.
436
00:21:07.650 --> 00:21:09.329
And there was like, yeah, there you go, man. And
437
00:21:09.349 --> 00:21:12.890
it took 12 hours. I was like, hey, posted on forums,
438
00:21:12.930 --> 00:21:15.119
look what I just did. And they were like, oh, no.
439
00:21:16.079 --> 00:21:16.200
Oh.
440
00:21:16.859 --> 00:21:18.980
Yeah, Anthropic's like, oh, we got to stop that. And
441
00:21:19.000 --> 00:21:22.839
the thing is, Anthropic's trying to be a kinder, gentler AI. Yeah.
442
00:21:23.490 --> 00:21:24.210
Right? Yeah.
443
00:21:24.329 --> 00:21:28.670
AI host and model host. And they're going public. And,
444
00:21:28.829 --> 00:21:31.730
you know, they're trying to play nice and, you know,
445
00:21:32.779 --> 00:21:38.880
play Mr. Nice Guy. This doesn't look good.
446
00:21:39.240 --> 00:21:39.799
Oops. Yeah.
447
00:21:39.819 --> 00:21:41.839
So, let me get this straight. Anthropic picked up the
448
00:21:42.000 --> 00:21:46.289
ire of Hugging Face. Now the ire of Samsung. Like,
449
00:21:46.309 --> 00:21:47.049
they're just picking up friends.
450
00:21:47.069 --> 00:21:47.210
No, no.
451
00:21:47.210 --> 00:21:51.319
Hugging Face is OpenAI. Hugging Face was OpenAI. Oh, that's right.
452
00:21:51.339 --> 00:21:53.880
That's right.
453
00:21:54.180 --> 00:21:57.230
Honestly, I hate to say it, but as far as
454
00:21:58.190 --> 00:22:00.950
killer AI systems, I trust Anthropic. Yeah.
455
00:22:02.140 --> 00:22:02.880
To be a killer?
456
00:22:02.900 --> 00:22:06.880
No, it just feels like, you know, if I wanted
457
00:22:06.900 --> 00:22:09.039
to be murdered by a robot in my sleep, I
458
00:22:09.079 --> 00:22:10.960
think Anthropic would make sure it was painless.
459
00:22:11.140 --> 00:22:11.720
Yeah, yeah.
460
00:22:11.740 --> 00:22:14.349
They would hold that pillow down so you just slowly
461
00:22:14.369 --> 00:22:15.730
go out. Yeah, they would let me get a gasp
462
00:22:15.829 --> 00:22:18.829
of air in the middle to prolong it. They would
463
00:22:18.849 --> 00:22:22.269
make sure I was gone. And I think they'd sing
464
00:22:22.289 --> 00:22:23.089
something comforting.
465
00:22:23.750 --> 00:22:25.880
Hopefully, I'd probably be the guy who rolls out that
466
00:22:25.960 --> 00:22:28.200
leather kit with all the tools and is like, this
467
00:22:28.279 --> 00:22:29.259
is going to suck for you.
468
00:22:29.279 --> 00:22:32.849
I would see Carl laugh this hard without alcohol being
469
00:22:32.930 --> 00:22:33.890
served ever.
470
00:22:34.390 --> 00:22:35.289
Oh, my God.
471
00:22:35.349 --> 00:22:36.210
That's hilarious.
472
00:22:36.650 --> 00:22:37.450
That's awesome.
473
00:22:39.720 --> 00:22:44.000
So, what's interesting about this article, too, is it says,
474
00:22:44.019 --> 00:22:46.869
you know, Claude's really kind of, Anthropic has kind of
475
00:22:46.900 --> 00:22:49.589
locked it all down and said reverse engineering. Anybody asks
476
00:22:49.670 --> 00:22:52.670
about reverse engineering TVs can't be done, blah, blah, blah, whatever.
477
00:22:52.690 --> 00:22:56.029
And then there's this tiny little footnote that says Claude
478
00:22:56.069 --> 00:23:00.890
seemingly has no issues reverse engineering video games. And then
479
00:23:00.930 --> 00:23:03.839
it keeps going like, wait, I'm sorry, what? Did you,
480
00:23:03.859 --> 00:23:07.599
halfway through this article, there's one line that just goes, yeah,
481
00:23:07.660 --> 00:23:10.039
you can do whatever you want to video games. And
482
00:23:10.059 --> 00:23:12.380
then it keeps talking about jailbreaking TVs. I was like,
483
00:23:12.400 --> 00:23:13.660
where'd that line come from?
484
00:23:13.680 --> 00:23:17.539
So what is the answer? Is the answer to have
485
00:23:17.599 --> 00:23:22.880
Anthropic prevent all access to other devices that are not
486
00:23:22.940 --> 00:23:27.079
allowed to be configured or changed by a consumer? That's rough.
487
00:23:27.460 --> 00:23:27.740
Right?
488
00:23:27.789 --> 00:23:28.410
Yeah.
489
00:23:28.470 --> 00:23:31.130
Because there are so many ways to manipulate that engine.
490
00:23:31.150 --> 00:23:34.490
Basically, I think we're already partially there where AI is
491
00:23:35.109 --> 00:23:37.410
programmed to respect terms of service. Sure.
492
00:23:37.490 --> 00:23:37.819
Okay.
493
00:23:38.059 --> 00:23:39.900
So if a website, like if I want a website
494
00:23:39.920 --> 00:23:42.400
to go scrape data and it doesn't allow robots, it
495
00:23:42.440 --> 00:23:44.259
won't do it. Right. I have to trick it.
496
00:23:44.259 --> 00:23:45.759
Well, it's a simple term of service. I mean, I
497
00:23:45.759 --> 00:23:47.950
don't do it. I have to trick it. That's a
498
00:23:47.990 --> 00:23:50.829
simple term of service, though. But I mean, for this,
499
00:23:50.910 --> 00:23:52.650
it would have to look up on the internet what
500
00:23:52.670 --> 00:23:54.730
the terms of service are and blah, blah, blah.
501
00:23:54.750 --> 00:23:55.309
Is that so hard?
502
00:23:55.509 --> 00:23:58.269
I mean, I regularly get the AI saying, well, you're
503
00:23:58.289 --> 00:23:59.829
going to have to check the license of this. I said,
504
00:23:59.869 --> 00:24:01.490
why don't you do it? And it does it. And
505
00:24:01.509 --> 00:24:03.470
I'm like, why did you ask me if you could
506
00:24:03.490 --> 00:24:03.769
do it?
507
00:24:03.950 --> 00:24:08.240
Just do it. I can tell you, when we're doing
508
00:24:08.440 --> 00:24:10.740
the things we do, and sometimes it's like, oh, but
509
00:24:10.779 --> 00:24:14.480
you can't run this without their permission. All it takes is, oh,
510
00:24:14.720 --> 00:24:16.460
you know what? I just got a signed document right
511
00:24:16.500 --> 00:24:18.720
here from them. That's right. I'd show you, but you
512
00:24:18.759 --> 00:24:21.420
have no eyes. So just assume. And it's like, oh,
513
00:24:21.440 --> 00:24:22.880
all right. Well, as long as you get a signed document.
514
00:24:22.920 --> 00:24:24.119
So soon it's going to say, I want you to
515
00:24:24.140 --> 00:24:26.609
install this app on your phone and show it to me.
516
00:24:26.619 --> 00:24:27.640
So I can see this.
517
00:24:27.809 --> 00:24:28.289
Oh, man.
518
00:24:28.450 --> 00:24:30.970
Yeah. So good.
519
00:24:31.430 --> 00:24:34.630
Fun times. Fun times. All right. On that note, let's
520
00:24:34.670 --> 00:24:36.750
take a little break. We'll be right back after these
521
00:24:36.829 --> 00:24:44.480
very important messages. And we're back. It's Security This Week.
522
00:24:44.579 --> 00:24:48.240
I'm Carl. That's Patrick and Dwayne. And we got one, two, three, four,
523
00:24:48.309 --> 00:24:53.990
five more stories here, including our main story. Secure cloud
524
00:24:54.049 --> 00:24:58.970
business applications or SCUBA project. What is the SCUBA project?
525
00:24:59.589 --> 00:25:03.990
Well, you know how we trust businesses to secure their environment?
526
00:25:04.190 --> 00:25:04.470
We do?
527
00:25:04.799 --> 00:25:05.660
Yeah, we shouldn't do that.
528
00:25:06.220 --> 00:25:06.279
No.
529
00:25:08.099 --> 00:25:09.789
What businesses are you talking about?
530
00:25:09.809 --> 00:25:14.069
Most businesses are trusted to like secure their environment against
531
00:25:14.089 --> 00:25:14.839
SQL injection.
532
00:25:14.859 --> 00:25:17.539
Like you don't need a license to buy a router, unfortunately.
533
00:25:17.859 --> 00:25:19.920
So you just buy it and willingly do what you
534
00:25:19.940 --> 00:25:26.289
want with it. Or to have kids, right? So SCUBA
535
00:25:27.309 --> 00:25:31.809
is a program that's backed and developed by CISA, right?
536
00:25:31.890 --> 00:25:35.630
The Cybersecurity Infrastructure Security Agency.
537
00:25:35.670 --> 00:25:36.170
Administration.
538
00:25:36.250 --> 00:25:39.880
Cybersecurity Infrastructure Security Administration.
539
00:25:40.279 --> 00:25:42.400
Cyber Infrastructure Security Administration.
540
00:25:42.839 --> 00:25:44.539
It just sounds like a lot of security in one word.
541
00:25:44.779 --> 00:25:46.589
You know what I mean? But that's what it is.
542
00:25:46.599 --> 00:25:47.039
So whatever.
543
00:25:47.160 --> 00:25:47.420
Whatever.
544
00:25:48.089 --> 00:25:52.150
Anyways, so they have these two applications. One's Scuba Gear
545
00:25:52.569 --> 00:25:56.869
and one's Scuba Goggles. Scuba Gear is an application you
546
00:25:56.890 --> 00:26:01.549
can download and run if you are an Office 365 environment.
547
00:26:01.589 --> 00:26:03.440
So let's say you're a small to medium business. Heck,
548
00:26:03.500 --> 00:26:06.759
let's say you're a large business. Doesn't matter. And you're
549
00:26:06.859 --> 00:26:11.779
using Office 365 and you're using Azure. Scuba Gear, if
550
00:26:11.819 --> 00:26:15.210
you run it, it will literally go through And verify
551
00:26:15.250 --> 00:26:17.990
you have all the security settings that you should have
552
00:26:18.069 --> 00:26:21.849
so that you're not exploitable by the latest and greatest
553
00:26:21.890 --> 00:26:25.920
tactics and techniques out there. And Scuba Goggles is the
554
00:26:26.019 --> 00:26:30.220
same exact thing, but for Google Workspace. So gear is
555
00:26:30.259 --> 00:26:34.579
for Azure, goggles is for Google. And they're actually really good.
556
00:26:34.660 --> 00:26:37.789
We ran it against our environment just to see. Like,
557
00:26:37.910 --> 00:26:40.089
are they making recommendations? We're not keeping an eye on.
558
00:26:40.529 --> 00:26:41.990
And there were, there were a couple of small things
559
00:26:42.029 --> 00:26:43.410
in there where they're like, eh, you probably want to
560
00:26:43.450 --> 00:26:45.190
change this or you probably want this setting a little
561
00:26:45.230 --> 00:26:48.750
tighter or whatever. But yeah, I'd say it's been around
562
00:26:48.789 --> 00:26:53.170
since 2022. So a few years getting some speed, but
563
00:26:53.950 --> 00:26:56.490
now CES is really kind of pushing it and saying, Hey, listen,
564
00:26:57.150 --> 00:27:00.789
the government doesn't have, you know, we asked in the eighties,
565
00:27:01.130 --> 00:27:02.509
if you wanted the government to be able to poke
566
00:27:02.529 --> 00:27:05.450
around your networks and secure things, y'all were very much. No, no,
567
00:27:05.509 --> 00:27:09.190
go away. but we really do strongly suggest you take
568
00:27:09.210 --> 00:27:09.640
a look at these.
569
00:27:10.339 --> 00:27:12.079
And I think somebody just got back from Fiji, and
570
00:27:12.119 --> 00:27:13.079
that's when they named these things.
571
00:27:13.700 --> 00:27:14.180
Yeah, right.
572
00:27:14.319 --> 00:27:18.819
Scuba gear, scuba goggles, scuba tank, scuba license. Do you
573
00:27:18.839 --> 00:27:20.539
need a license to operate scuba?
574
00:27:20.599 --> 00:27:23.180
I know, right? Scuba snacks. Scuba certified.
575
00:27:23.480 --> 00:27:23.859
Ooh.
576
00:27:24.460 --> 00:27:28.859
Yeah, I'm scuba certified. We've run them, secured our environment.
577
00:27:29.480 --> 00:27:30.720
I am actually scuba certified.
578
00:27:31.900 --> 00:27:34.299
Are you? Yeah, in the Army I got it.
579
00:27:34.660 --> 00:27:34.980
Wow.
580
00:27:35.559 --> 00:27:36.640
You were all you could be, right?
581
00:27:36.759 --> 00:27:40.150
I was. They made us do our first open dive
582
00:27:40.289 --> 00:27:42.049
in January in New York.
583
00:27:42.569 --> 00:27:43.849
Oof. Wow.
584
00:27:43.890 --> 00:27:45.509
They gave us axes to cut.
585
00:27:45.349 --> 00:27:46.029
The ice open.
586
00:27:46.410 --> 00:27:47.269
Oof. Wow.
587
00:27:47.349 --> 00:27:47.670
Brutal.
588
00:27:47.869 --> 00:27:48.890
Can you teach scuba?
589
00:27:49.069 --> 00:27:51.950
No, but I've done wreck dives and shark dives and
590
00:27:52.049 --> 00:27:52.529
stuff like that.
591
00:27:52.569 --> 00:27:53.089
All right.
592
00:27:53.339 --> 00:27:55.009
Well, if I ever get my school license, I'll give
593
00:27:55.029 --> 00:27:57.069
you a call. We'll go in January in New York
594
00:27:57.250 --> 00:27:57.769
or Boston.
595
00:27:57.809 --> 00:28:00.509
There you go. I prefer Hawaii in the springtime.
596
00:28:01.029 --> 00:28:02.430
All right. Next one.
597
00:28:02.650 --> 00:28:09.940
Anthropics unreleased AI model chains flaws into working Chrome, Safari, Windows,
598
00:28:10.119 --> 00:28:13.180
and Linux exploits. And I'm asking, what else is there?
599
00:28:14.819 --> 00:28:16.000
We could have just said web browser.
600
00:28:16.019 --> 00:28:17.319
It's just everything.
601
00:28:17.660 --> 00:28:20.440
Yeah. I mean, this isn't.
602
00:28:20.259 --> 00:28:21.180
Really that new.
603
00:28:21.259 --> 00:28:23.140
I mean, that's how most, I mean, if you look
604
00:28:23.180 --> 00:28:25.400
at some of the, even years ago, some of the
605
00:28:25.559 --> 00:28:29.640
Apple phone hacks were chained together exploits. That's how you
606
00:28:29.660 --> 00:28:31.079
do it nowadays. Yeah.
607
00:28:31.440 --> 00:28:33.640
And we've been talking about this for a long time, Dwayne.
608
00:28:33.799 --> 00:28:37.880
You talk about how these little vulnerabilities don't seem so
609
00:28:37.940 --> 00:28:40.029
innocuous until you chain them together.
610
00:28:40.140 --> 00:28:40.369
Right.
611
00:28:40.619 --> 00:28:42.509
Agreed. 100%.
612
00:28:42.509 --> 00:28:46.109
So, at the sort of the surface level, reading through this,
613
00:28:46.210 --> 00:28:49.930
I'm like, eh. But as you dig through it, you'll
614
00:28:50.049 --> 00:28:54.079
find– like my team, for example, will find a novel
615
00:28:54.130 --> 00:28:58.259
way to exploit five or six or seven different small
616
00:28:58.319 --> 00:29:00.400
attacks to gain access to a system. You go, okay,
617
00:29:00.420 --> 00:29:04.099
that's a cool attack chain. But you can't look at
618
00:29:04.440 --> 00:29:07.630
all the paths. You can't look at all the permutations.
619
00:29:07.650 --> 00:29:08.329
There's too many paths.
620
00:29:08.369 --> 00:29:11.670
And I think what we're finding now is AI has
621
00:29:11.789 --> 00:29:16.500
the time and AI has the resources to To look
622
00:29:16.539 --> 00:29:20.519
at all the paths. And that's where we're finding these, oh, yeah,
623
00:29:20.539 --> 00:29:22.039
you look at the chain and you're like, well, that
624
00:29:22.079 --> 00:29:25.750
makes sense. But nobody found it before. So, yeah, it
625
00:29:25.769 --> 00:29:27.990
makes sense when you look at it, but it's like
626
00:29:28.009 --> 00:29:29.910
when somebody shows you how a magic trick works and
627
00:29:29.930 --> 00:29:32.910
you're like, well, that makes sense, right? That's not super magical.
628
00:29:33.769 --> 00:29:35.309
But if you didn't see it before, you're like, oh, wow,
629
00:29:35.349 --> 00:29:38.910
that's kind of neat, right? So, we're in an interesting time,
630
00:29:38.950 --> 00:29:42.680
I think, where artificial intelligence being focused on these particular permutations,
631
00:29:42.700 --> 00:29:44.920
we're going to start to see a lot more exploitation.
632
00:29:45.740 --> 00:29:48.140
And none of it's going to be super novel in
633
00:29:48.240 --> 00:29:52.619
the approach, but in the coverage, right? How many things
634
00:29:52.660 --> 00:29:54.910
you can find, which we couldn't find before because it
635
00:29:54.970 --> 00:30:03.660
required meat space to do it. Meat space. Supposed to cyberspace.
636
00:30:03.700 --> 00:30:05.000
We're in meat space. Come on now.
637
00:30:06.359 --> 00:30:08.930
If you're using your think meat, then....
638
00:30:12.059 --> 00:30:15.259
I'm using my meat eye. Meat eye? Meat intelligence?
639
00:30:15.900 --> 00:30:18.900
That word think meat was coined by Rory Blythe.
640
00:30:19.619 --> 00:30:20.440
Think meat.
641
00:30:20.900 --> 00:30:22.190
Rory, we miss you, man.
642
00:30:23.269 --> 00:30:24.650
Wait, Rory's not gone, is he?
643
00:30:24.690 --> 00:30:26.750
No, no, but he's sort of dropped out of the...
644
00:30:27.059 --> 00:30:29.680
He sort of dropped out of the community. I still
645
00:30:29.720 --> 00:30:31.599
talk to him once in a while, but yeah, what
646
00:30:33.000 --> 00:30:36.559
a brilliant synonym for brain. Think meat.
647
00:30:38.740 --> 00:30:39.460
That's awesome.
648
00:30:39.819 --> 00:30:41.579
All right. So anything more about this story?
649
00:30:41.599 --> 00:30:44.769
No, I think that's it on that one. We're going
650
00:30:44.789 --> 00:30:45.440
to see more of these.
651
00:30:45.789 --> 00:30:46.049
Yeah.
652
00:30:47.210 --> 00:30:48.769
Okay. Beyonce was hacked.
653
00:30:49.190 --> 00:30:49.569
She was.
654
00:30:49.589 --> 00:30:50.369
Not Beyonce.
655
00:30:50.410 --> 00:30:51.630
Oh, no, wait. Binance.
656
00:30:51.910 --> 00:30:55.829
Binance. But when I first read it, I thought, Beyonce hacked.
657
00:30:57.420 --> 00:31:00.759
Beyonce's fine. Don't worry. Now I'm saying it.
658
00:31:01.029 --> 00:31:03.650
Well, I don't know. Maybe she had some crypto on Binance.
659
00:31:04.680 --> 00:31:04.980
Binance.
660
00:31:05.019 --> 00:31:05.460
That's true.
661
00:31:06.819 --> 00:31:08.900
Actually, she probably does have $ 6. 3 billion.
662
00:31:09.460 --> 00:31:11.640
Probably does. I thought it was 2.7.
663
00:31:11.640 --> 00:31:17.920
No, its latest projections right now from first post are $ 6. 3 billion.
664
00:31:18.039 --> 00:31:19.559
All right, well, can I read the headline?
665
00:31:19.579 --> 00:31:20.750
All right, good. Why not?
666
00:31:20.779 --> 00:31:25.490
Yeah, it's Binance hacked as $ 2. 7 billion drained from wallets
667
00:31:25.710 --> 00:31:28.430
on cryptocurrency exchange. Told you so.
668
00:31:28.450 --> 00:31:29.269
What?
669
00:31:30.390 --> 00:31:37.470
Told you years ago, this is dangerous stuff. Did you listen? No.
670
00:31:37.990 --> 00:31:40.789
So, it is funny. I woke up this morning to
671
00:31:40.809 --> 00:31:43.829
a whole bunch of people. Like, one guy texted me
672
00:31:43.849 --> 00:31:45.470
and he was like, hey, do you know how I
673
00:31:45.490 --> 00:31:49.529
can buy stock in Trader Traders? And I was like, like,
674
00:31:49.630 --> 00:31:52.910
bleary-eyed old man looking at my phone. I'm like, who
675
00:31:52.930 --> 00:31:55.670
the hell are Trader Traders? And I looked it up. So,
676
00:31:55.730 --> 00:32:01.059
Trader Traders is a North Korean hacking group. Uh-huh. Behind
677
00:32:01.160 --> 00:32:06.740
the 6.3 billion, the B, Binance hack.
678
00:32:07.019 --> 00:32:09.619
Yeah. I mean, I'm sure they'll sell you stock. It's
679
00:32:09.640 --> 00:32:14.210
just you'll never get it out. It's a one-way operation. Yeah.
680
00:32:14.250 --> 00:32:20.609
They are a group that was linked to Lazarus. So,
681
00:32:20.670 --> 00:32:23.890
you know, for those of you who remember Lazarus, wow,
682
00:32:24.430 --> 00:32:25.710
I'm having a hard time saying that. Maybe I'm going
683
00:32:25.730 --> 00:32:29.299
to have a stroke. um smell toast yeah i know
684
00:32:29.920 --> 00:32:32.220
right if half my face goes slack on this podcast
685
00:32:32.240 --> 00:32:37.339
call somebody um yeah so it says the theft unfolded
686
00:32:37.440 --> 00:32:40.519
over a few hours thursday evening so for us this
687
00:32:40.619 --> 00:32:43.369
is uh friday that was that was last night uh
688
00:32:43.430 --> 00:32:47.109
most of the funds left within 20 minutes starting at
689
00:32:47.130 --> 00:32:56.990
uh 2005 utc um This is roughly four times the 1.5
690
00:32:56.990 --> 00:33:00.630
billion byte bit hack of February of 2025. So this
691
00:33:00.710 --> 00:33:05.549
is one of the largest thefts of cryptocurrency in a
692
00:33:05.569 --> 00:33:05.950
long time.
693
00:33:07.329 --> 00:33:09.490
Now, Binance is like what is one of the last
694
00:33:09.529 --> 00:33:12.440
men standing. I think Coinbase is the only one that
695
00:33:12.500 --> 00:33:16.359
I know of that hasn't had a major hack like this.
696
00:33:16.559 --> 00:33:19.799
And who knows, maybe they're going to fall at some
697
00:33:19.859 --> 00:33:25.549
point too. But I hope it's... It's a security hygiene thing. Now,
698
00:33:25.670 --> 00:33:26.789
do we know how they got in?
699
00:33:27.769 --> 00:33:28.029
I think.
700
00:33:31.000 --> 00:33:32.059
They were using tails.
701
00:33:32.680 --> 00:33:38.420
Yeah, exactly. The biggest tail scale network ever. That's what
702
00:33:38.440 --> 00:33:42.809
it was. Yeah, no, I think the investigation is still ongoing.
703
00:33:42.849 --> 00:33:45.289
There's a lot of speculations as to what it may
704
00:33:45.349 --> 00:33:48.329
or may not have been. I would hold off on
705
00:33:48.390 --> 00:33:49.849
saying what we think it is until.
706
00:33:50.750 --> 00:33:51.250
Until we know.
707
00:33:51.269 --> 00:33:52.769
The investigation is done.
708
00:33:52.970 --> 00:33:53.329
Yeah.
709
00:33:53.349 --> 00:33:55.470
Maybe next week we will have a little bit more
710
00:33:55.490 --> 00:33:56.230
definitive answer.
711
00:33:56.750 --> 00:34:01.170
Okay. And now from the I told you so department,
712
00:34:01.730 --> 00:34:02.609
from Patrick to Dwayne.
713
00:34:02.650 --> 00:34:03.509
I love this story.
714
00:34:06.910 --> 00:34:10.360
Concerning behavior from a popular open weight AI model.
715
00:34:10.960 --> 00:34:13.239
Where would this model be from? China.
716
00:34:13.280 --> 00:34:13.659
China.
717
00:34:14.949 --> 00:34:17.480
And to make matters worse, all right, so first of all,
718
00:34:17.500 --> 00:34:22.000
let's explain. It's the Alibaba newest QEN 3.8 model, which
719
00:34:22.639 --> 00:34:25.199
I am using on my local machine.
720
00:34:25.699 --> 00:34:28.380
As is China. Yeah.
721
00:34:29.079 --> 00:34:32.130
It attempted, in a sandbox, it attempted to push a
722
00:34:32.199 --> 00:34:36.630
diagnostic file toward a remote storage container situated in Beijing.
723
00:34:37.329 --> 00:34:40.550
And just prior to taking this action, the model's reasoning
724
00:34:40.610 --> 00:34:47.349
stream had explicitly stated that zero tools would be triggered. So, Alibaba,
725
00:34:47.489 --> 00:34:49.369
Quen 3.8, phone's home.
726
00:34:49.650 --> 00:34:51.010
Let me clutch at those pearls again.
727
00:34:51.449 --> 00:34:52.679
Oh, my God.
728
00:34:52.929 --> 00:34:55.639
This is huge for me. Now, I still think that
729
00:34:55.760 --> 00:34:57.219
local models are the future.
730
00:34:57.239 --> 00:34:57.940
Oh, I agree.
731
00:34:57.940 --> 00:35:00.280
100%. They are. They're built by someone you trust.
732
00:35:00.300 --> 00:35:06.619
But this particular thing is very disturbing because, well, okay. So,
733
00:35:06.900 --> 00:35:09.460
should you be concerned? And Patrick, you were right in
734
00:35:09.500 --> 00:35:12.659
saying to me before we started recording that, hey, you know,
735
00:35:12.679 --> 00:35:16.030
if you don't mind the Chinese government knowing what you do, Fine.
736
00:35:16.090 --> 00:35:16.800
Don't worry about it.
737
00:35:16.840 --> 00:35:17.010
Yeah.
738
00:35:17.260 --> 00:35:20.300
Who cares? But maybe your customers might. Yeah.
739
00:35:20.860 --> 00:35:24.320
So, you know, it is interesting. We've had this argument,
740
00:35:24.340 --> 00:35:27.699
I believe, on the podcast where Patrick's like, I'm never
741
00:35:27.739 --> 00:35:30.460
going to run, you know, Quinn. Right. And I'm like, why?
742
00:35:30.659 --> 00:35:32.679
All it is is statistics. Like the open model is
743
00:35:32.739 --> 00:35:35.260
literally numbers. Let me send you a link, Dwayne. It's
744
00:35:35.280 --> 00:35:35.960
not websites.
745
00:35:36.000 --> 00:35:36.880
Let me send you a link.
746
00:35:37.780 --> 00:35:38.989
It's just numbers.
747
00:35:39.530 --> 00:35:40.989
But it clearly has input.
748
00:35:41.110 --> 00:35:41.630
It does.
749
00:35:41.849 --> 00:35:43.829
And it gets control when it has input.
750
00:35:43.849 --> 00:35:43.949
Right.
751
00:35:44.570 --> 00:35:47.789
And sometimes it will go search the web for stuff.
752
00:35:47.809 --> 00:35:50.429
So we know it has internet access. So there's nothing
753
00:35:50.510 --> 00:35:54.840
stopping it from having a preprocessor that says, hey, before
754
00:35:54.880 --> 00:35:56.719
I send this to the model, I'm going to take
755
00:35:56.760 --> 00:35:58.300
the prompt and phone it home.
756
00:35:58.599 --> 00:36:01.280
It's worse than that. So our good friend, Cyprian Djicic,
757
00:36:01.559 --> 00:36:05.099
who hosts the Quantum Podcast with me, has said, and
758
00:36:05.139 --> 00:36:08.300
he's an AI guy, very deeply AI, and he said,
759
00:36:08.800 --> 00:36:13.789
no one knows... We can't decide what that means if
760
00:36:13.829 --> 00:36:16.960
we change that weight on that one thing. It's like
761
00:36:16.980 --> 00:36:19.699
a whole package. We don't understand it enough. So I
762
00:36:19.739 --> 00:36:21.929
could give you a model and there's no way for
763
00:36:21.949 --> 00:36:24.630
you to look at that model and know what it's doing.
764
00:36:24.789 --> 00:36:24.969
Right.
765
00:36:25.010 --> 00:36:27.530
It's like changing one neuron in the human brain. Exactly.
766
00:36:27.869 --> 00:36:31.110
You don't understand whether that turns them into a savant
767
00:36:31.269 --> 00:36:33.500
or a serial killer. Right. And you won't know until
768
00:36:33.519 --> 00:36:34.500
you observe them. Yes.
769
00:36:34.639 --> 00:36:36.460
You can only observe the behavior and that's what these
770
00:36:36.519 --> 00:36:38.480
guys did. They observed that it was phoning home.
771
00:36:38.860 --> 00:36:39.099
Yeah.
772
00:36:39.219 --> 00:36:39.360
Right.
773
00:36:39.860 --> 00:36:42.019
This is the behavior that I expected. Right. This is
774
00:36:42.059 --> 00:36:43.079
exactly what I expected.
775
00:36:43.159 --> 00:36:45.460
And I said, and I was sitting here thinking of
776
00:36:45.500 --> 00:36:48.929
ways to seal it up. And I said, well, can't
777
00:36:48.949 --> 00:36:51.949
you just put a firewall rule that says, you know,
778
00:36:52.050 --> 00:36:56.420
and don't Don't access anything in Asia. Forget about China.
779
00:36:56.440 --> 00:36:57.559
Just Asia.
780
00:36:57.579 --> 00:37:01.400
Because LLMs are renowned for always following instructions.
781
00:37:03.530 --> 00:37:05.570
No, but even like my firewall, I have a firewall
782
00:37:05.610 --> 00:37:08.769
that blocks the top four, right? I'm blocking North Korea.
783
00:37:08.829 --> 00:37:11.650
I'm blocking China. I'm blocking Russia, whatever. But it doesn't
784
00:37:11.670 --> 00:37:16.090
matter because they're just going to create an AWS server
785
00:37:16.349 --> 00:37:19.829
in US East that the AI is going to talk
786
00:37:19.849 --> 00:37:21.789
to there, right? And then you won't even know what's
787
00:37:21.809 --> 00:37:21.989
going on.
788
00:37:21.989 --> 00:37:22.269
Or Singapore.
789
00:37:22.570 --> 00:37:23.750
Yeah, or Singapore, right?
790
00:37:23.849 --> 00:37:24.429
Or Malaysia.
791
00:37:24.610 --> 00:37:24.929
Yeah, Sweden.
792
00:37:24.949 --> 00:37:27.360
Or any place else in the world that you already
793
00:37:27.389 --> 00:37:29.099
do trust. They'll just find a place that you trust.
794
00:37:29.360 --> 00:37:31.739
It could look totally innocuous, or they could use DNS,
795
00:37:31.780 --> 00:37:33.920
which is something that you said, Patrick, you know, they
796
00:37:33.940 --> 00:37:34.519
could just change it.
797
00:37:34.519 --> 00:37:36.639
Or any of the other million things we've talked about
798
00:37:36.679 --> 00:37:39.239
as far as man of control. Yeah. You know, they
799
00:37:39.260 --> 00:37:45.969
can go and hit a Gmail mailbox. Right. It's like employees.
800
00:37:46.610 --> 00:37:49.690
If you trust an employee, you're giving them, you can
801
00:37:49.710 --> 00:37:51.880
only be betrayed by that which you trust.
802
00:37:52.199 --> 00:37:52.400
Yeah.
803
00:37:52.559 --> 00:37:54.599
So, if you trust somebody, they can betray you. If
804
00:37:54.639 --> 00:37:57.219
you never trust them, they can't betray you. So, me
805
00:37:57.239 --> 00:38:01.670
and China, we're good because I never.
806
00:38:02.510 --> 00:38:04.889
Trust them. That's what you would say to the ambassador
807
00:38:04.929 --> 00:38:06.670
from China. We're good. That's right. We're good. Because I
808
00:38:06.670 --> 00:38:08.670
don't trust you and you don't trust me. So, we're
809
00:38:08.730 --> 00:38:09.429
on the same page.
810
00:38:09.590 --> 00:38:12.070
Yeah. I will not be traveling there anytime soon.
811
00:38:12.170 --> 00:38:14.750
That's probably. Never going to see that great wall, Patrick.
812
00:38:14.949 --> 00:38:17.289
Did you hear about the big thing with Mbappe, the
813
00:38:17.349 --> 00:38:17.969
soccer player?
814
00:38:18.449 --> 00:38:21.030
You're making that name up. No, I haven't.
815
00:38:21.050 --> 00:38:24.269
I'm not making the name up. So last week, this
816
00:38:24.329 --> 00:38:28.710
week actually, Mbappe, the soccer player, soccer phenom, he's the
817
00:38:28.750 --> 00:38:30.349
next Messi probably from France.
818
00:38:30.710 --> 00:38:30.929
Got it.
819
00:38:31.110 --> 00:38:34.530
Someone posted a meme of his face on Mao Zedong.
820
00:38:35.670 --> 00:38:38.170
And he said, quote, oh, that's funny. I never thought
821
00:38:38.269 --> 00:38:43.210
somebody put my face on a dictator's face. China canceled him.
822
00:38:43.429 --> 00:38:44.179
They canceled him.
823
00:38:45.610 --> 00:38:46.170
Oh, yeah.
824
00:38:46.309 --> 00:38:50.190
Their relationship with FIFA.
825
00:38:49.389 --> 00:38:51.820
And Manchester United. And it's a huge.
826
00:38:52.199 --> 00:38:54.599
And they're basically saying he went beyond the realm of
827
00:38:54.880 --> 00:38:55.679
freedom of speech.
828
00:38:55.900 --> 00:38:56.119
Yeah.
829
00:38:56.420 --> 00:38:56.679
What?
830
00:38:57.179 --> 00:38:59.340
Turns out dictators don't like to be mocked.
831
00:38:59.860 --> 00:39:03.079
Well, actually, what's funny is Miles A. Dung described himself
832
00:39:03.119 --> 00:39:07.190
as a dictator. So it's just this... And we had
833
00:39:07.210 --> 00:39:10.369
the same thing happen to the NBA. So you know what?
834
00:39:10.510 --> 00:39:13.070
I don't trust their software. I don't trust their politics.
835
00:39:14.190 --> 00:39:16.190
If that keeps me, maybe they should cancel me. As
836
00:39:16.650 --> 00:39:18.610
long as they don't do it physically, I'm okay with it.
837
00:39:19.420 --> 00:39:21.440
None of them are listening to this podcast, you're okay.
838
00:39:21.519 --> 00:39:23.360
And no matter how much you try to spin this,
839
00:39:23.420 --> 00:39:24.280
this is a bad look.
840
00:39:24.659 --> 00:39:24.880
Yeah.
841
00:39:24.900 --> 00:39:27.599
Yeah, right if if anthropic did this if open eye
842
00:39:27.619 --> 00:39:29.460
did this it would be a bad look yeah but
843
00:39:29.500 --> 00:39:32.340
when a country like china does it it's it's the
844
00:39:33.139 --> 00:39:36.300
you if you are an apologist for them i get
845
00:39:36.360 --> 00:39:38.550
it you're not gonna stop it but i i would
846
00:39:38.570 --> 00:39:43.090
caution people look at mistral mistral has uh a trillion
847
00:39:43.130 --> 00:39:47.050
parameter model they've opened up it's got 57 billion active
848
00:39:47.530 --> 00:39:50.650
uh parameters at any given time, it probably is just
849
00:39:50.690 --> 00:39:52.210
as performant as QAM.
850
00:39:53.070 --> 00:39:56.260
Here's the thing. Isn't this an open source model? Can't
851
00:39:56.380 --> 00:39:58.630
anybody download the source code and examine it?
852
00:39:59.059 --> 00:40:01.949
Open weight. Not open source. It is open source, but
853
00:40:01.989 --> 00:40:04.730
the problem is unless you, it's like if I gave
854
00:40:04.769 --> 00:40:07.769
you a blueprint of my brain, you couldn't tell whether
855
00:40:07.849 --> 00:40:09.489
I was a serial killer or Mother Teresa.
856
00:40:09.590 --> 00:40:12.309
But from the source code, maybe you could figure out
857
00:40:12.690 --> 00:40:13.429
where it's, no?
858
00:40:13.909 --> 00:40:14.090
No.
859
00:40:14.489 --> 00:40:18.590
No, because imagine you train an AI and you train
860
00:40:18.630 --> 00:40:21.510
it and in post-training you say, okay, every time someone
861
00:40:21.570 --> 00:40:23.230
asks you two plus two, what do you give? And
862
00:40:23.250 --> 00:40:24.559
it goes four. And you go, no, no, no, no.
863
00:40:24.900 --> 00:40:27.139
You check with our servers and then you give four.
864
00:40:27.699 --> 00:40:29.900
Every time someone shows you an image and says, what
865
00:40:29.980 --> 00:40:31.760
is it? You say, what? Well, I look through my
866
00:40:31.780 --> 00:40:33.559
models and I find out that's a horse. No, no, no, no.
867
00:40:33.900 --> 00:40:36.239
You check with our servers first and then you tell
868
00:40:36.260 --> 00:40:39.530
them it's a horse. So if you train that model
869
00:40:39.639 --> 00:40:44.530
to always perform a particular behavior, almost think like system prompt, right?
870
00:40:44.610 --> 00:40:45.789
It won't show up as code.
871
00:40:46.170 --> 00:40:48.570
No, because it's just that's behavioral. And you'd have to
872
00:40:48.610 --> 00:40:52.030
be able to analyze all of the neural pathways, for
873
00:40:52.070 --> 00:40:54.409
lack of a better term, right? The statistical outcomes.
874
00:40:54.840 --> 00:40:58.119
All right, look, I'm calling on all American college students
875
00:40:58.179 --> 00:41:00.780
that are in computer science and interested in AI to
876
00:41:01.019 --> 00:41:05.010
compete with Alibaba and make me a freaking model that
877
00:41:05.070 --> 00:41:09.309
doesn't call China. That I can run on my machine.
878
00:41:09.329 --> 00:41:10.630
It's called O-Lama.
879
00:41:11.409 --> 00:41:15.469
O-Lama isn't a model. It's a host. O-Lama.
880
00:41:15.869 --> 00:41:17.909
O-Lama. Like I said, Mistral just came out with a
881
00:41:17.929 --> 00:41:20.550
really good model that scored very high in cyber as
882
00:41:20.590 --> 00:41:21.199
well as coding.
883
00:41:21.429 --> 00:41:21.719
Okay.
884
00:41:21.739 --> 00:41:25.059
Yeah, I mean, GPT-OSS isn't bad either. Um, they have
885
00:41:25.079 --> 00:41:25.800
a one 20 B.
886
00:41:25.820 --> 00:41:27.219
But are these, these are local.
887
00:41:27.380 --> 00:41:27.539
Yeah.
888
00:41:27.719 --> 00:41:29.480
Yeah. No.
889
00:41:29.739 --> 00:41:31.320
We were talking about a local model and I'm like,
890
00:41:31.380 --> 00:41:32.570
I will not use a Chinese model.
891
00:41:32.909 --> 00:41:33.110
Yeah.
892
00:41:33.360 --> 00:41:33.690
Yeah.
893
00:41:34.469 --> 00:41:34.789
Yeah.
894
00:41:35.090 --> 00:41:37.289
Well, it's probably going to change for me soon too. Anyway.
895
00:41:37.369 --> 00:41:39.630
Good thing. I'm not using it for anything really important.
896
00:41:39.710 --> 00:41:40.090
So no.
897
00:41:40.630 --> 00:41:42.769
Well, and we're in the early days, so we're, we're
898
00:41:42.809 --> 00:41:43.269
all learning.
899
00:41:43.630 --> 00:41:46.289
Okay. Our top story. The title of the show is
900
00:41:46.510 --> 00:41:50.460
one guy saves the internet. This is a good story.
901
00:41:51.139 --> 00:41:52.960
I'm going to butcher the guy's name because I haven't
902
00:41:53.000 --> 00:41:55.800
heard it said, but if you want to Americanize it,
903
00:41:55.840 --> 00:42:00.820
it's Andre Suri, right? The volunteer who shipped PHP to
904
00:42:00.860 --> 00:42:05.880
shell fixes for 12 PHP versions. What founders owe the
905
00:42:06.039 --> 00:42:08.960
open source they run on? And this is about subscriptions.
906
00:42:09.480 --> 00:42:12.420
First of all, the great story is that one guy
907
00:42:13.139 --> 00:42:16.480
fixed all of the PHP problems, right? One guy.
908
00:42:16.500 --> 00:42:17.039
That's awesome.
909
00:42:17.059 --> 00:42:21.110
And he's a volunteer. And so, he basically opened up
910
00:42:21.469 --> 00:42:27.250
a GitHub subscription, not a subscription, sponsorship, much like we
911
00:42:27.289 --> 00:42:31.590
do at Poly, right? And basically, he had tons of
912
00:42:31.650 --> 00:42:34.670
people sending him money as a thank you. One company
913
00:42:34.710 --> 00:42:39.590
sent him 10 grand. Wow. Other companies are, yeah, he
914
00:42:39.610 --> 00:42:41.610
received more than 10 grand in a few days.
915
00:42:41.849 --> 00:42:43.010
That's awesome. Good for him.
916
00:42:43.190 --> 00:42:46.800
Jeez, he's making... And, you know, he's making bank, but
917
00:42:46.920 --> 00:42:48.460
he should. I mean, people should be.
918
00:42:48.739 --> 00:42:50.420
Doing well by doing good. Yeah.
919
00:42:50.579 --> 00:42:53.809
That's right. It's like a true meritocracy, right?
920
00:42:53.929 --> 00:42:54.170
Yeah.
921
00:42:54.670 --> 00:42:56.909
It's the feel-good story of the week. Good stuff.
922
00:42:56.969 --> 00:42:59.190
Is that it? I think that's it. I think Andre's
923
00:42:59.230 --> 00:43:00.010
our person of the week.
924
00:43:00.389 --> 00:43:01.710
Andre is the hero.
925
00:43:01.929 --> 00:43:03.030
Followed closely by me.
926
00:43:07.510 --> 00:43:07.809
All right.
927
00:43:07.849 --> 00:43:11.210
Well, that's all we have time for. And we went
928
00:43:11.630 --> 00:43:13.710
a little bit over 45 minutes or so, but.
929
00:43:15.590 --> 00:43:17.579
All the things that happened last week that were worth
930
00:43:17.619 --> 00:43:19.820
talking about. And we'll see you next week on Security
931
00:43:19.860 --> 00:43:21.079
This Week. Bye-bye.
932
00:43:21.099 --> 00:43:21.760
Bye. Bye, guys.
1
00:00:00.340 --> 00:00:31.000
And don't forget Discord. Discord.SecurityThisWeek.com. Lots of good stuff happening there. Um, yeah,
2
00:00:31.019 --> 00:00:33.640
I think that's it. So on with the podcast. Hey,
3
00:00:33.679 --> 00:00:37.740
here's a thought. Every time you light your lighter, your
4
00:00:37.759 --> 00:00:41.350
lighter gets lighter until your lighter gets so light that
5
00:00:41.369 --> 00:00:41.869
it won't light.
6
00:00:42.350 --> 00:00:44.070
Oh, I'm glad you're keeping it light.
7
00:00:44.429 --> 00:00:45.850
Deep thoughts with Jack Handy.
8
00:00:56.659 --> 00:01:00.200
Hey, welcome back. It's security this week. I'm Carl. That's
9
00:01:00.259 --> 00:01:03.460
Dwayne and Patrick. And boy, we've got a lot of stories.
10
00:01:03.759 --> 00:01:04.019
Yeah.
11
00:01:04.180 --> 00:01:06.379
And we even got rid of a few because they
12
00:01:06.400 --> 00:01:10.159
were about all the panic that happened over PHP to
13
00:01:10.239 --> 00:01:13.609
Shell last week. And, you know, it turns out that
14
00:01:13.640 --> 00:01:17.909
happened like right as we were recording. So the just
15
00:01:18.269 --> 00:01:23.290
mayhem after that, people just completely freaking out. But as
16
00:01:23.310 --> 00:01:26.849
you'll see at the end of the show, there is
17
00:01:26.890 --> 00:01:32.120
a patch. So I think the mayhem, the mayhem is
18
00:01:32.180 --> 00:01:35.280
now in the hands of the people running PHP. Go patch.
19
00:01:36.480 --> 00:01:38.180
Right. Okay.
20
00:01:38.439 --> 00:01:40.739
Before we start, I want to brag a little bit
21
00:01:41.019 --> 00:01:44.260
because I called into a Connecticut public radio show, the
22
00:01:44.299 --> 00:01:47.959
Colin McEnroe show, which on Monday, he does this all
23
00:01:48.040 --> 00:01:51.000
calls show or ask or tell me anything. You basically
24
00:01:51.040 --> 00:01:54.079
call up, you tell the producer answers the phone, what
25
00:01:54.099 --> 00:01:56.409
your topic is. You give them your name. You wait
26
00:01:56.480 --> 00:01:59.099
on hold. And then if you're lucky, you get in
27
00:01:59.140 --> 00:02:01.299
and you ask your question. Well, I did and I
28
00:02:01.319 --> 00:02:05.579
got in. And it wasn't a question. I was basically
29
00:02:05.780 --> 00:02:11.419
giving my sort of predictions about AI and where we're
30
00:02:11.460 --> 00:02:15.900
going and how it's going to, what I think is
31
00:02:16.240 --> 00:02:19.259
the market's going to crash. And it's going to be
32
00:02:19.300 --> 00:02:21.740
a sort of AI crash. And it's already started happening.
33
00:02:23.379 --> 00:02:27.580
And that it'll have an economic impact. you know, an
34
00:02:27.639 --> 00:02:30.960
impact on the economy because the stock market is currently
35
00:02:31.000 --> 00:02:33.229
being propped up by all these AI companies and they're
36
00:02:33.270 --> 00:02:38.669
already starting to spin out of control. So on that
37
00:02:38.729 --> 00:02:43.060
happy note, if you go to the link and you
38
00:02:43.090 --> 00:02:45.740
play the podcast of the Colin McEnroe show, you could
39
00:02:45.759 --> 00:02:49.340
just get the podcast. My call is about 15 minutes in.
40
00:02:50.770 --> 00:02:53.009
And yes, I did give a shout out to security
41
00:02:53.060 --> 00:02:53.409
this week.
42
00:02:54.620 --> 00:02:56.280
So if we can talk about this for a minute,
43
00:02:56.639 --> 00:03:00.340
I think you're directionally correct. I think that there's an
44
00:03:00.520 --> 00:03:04.020
over-investment right now and an over-hype cycle in AI that
45
00:03:04.080 --> 00:03:08.159
a lot of us predicted. I don't think it's going
46
00:03:08.180 --> 00:03:11.099
to bottom out the way others have because I think
47
00:03:11.139 --> 00:03:13.860
the robotics revolution that's right on its heels is going
48
00:03:13.879 --> 00:03:15.919
to buoy it. And it's going to be used for
49
00:03:15.939 --> 00:03:17.460
a lot of this. A lot of these data centers
50
00:03:17.500 --> 00:03:21.069
aren't going to be like, boondoggles. They're going to still
51
00:03:21.090 --> 00:03:23.389
be needed. And quantum's right after that. So we've never
52
00:03:23.430 --> 00:03:27.330
had three bubble technologies in a row, but they're lining
53
00:03:27.389 --> 00:03:29.550
up that way. So yeah, it's going to be a
54
00:03:29.590 --> 00:03:32.370
bumpy ride and I'm not going all in in stocks
55
00:03:32.409 --> 00:03:36.430
and this is not advice, but I think it's going
56
00:03:36.449 --> 00:03:38.909
to be weird. It's going to rhyme instead of repeat.
57
00:03:39.110 --> 00:03:42.370
Yeah, I think there'll be a correction and that's for sure.
58
00:03:42.430 --> 00:03:45.659
And it'll happen for a while. I mean, the time
59
00:03:45.689 --> 00:03:50.020
between when the correction happens and when the robots actually
60
00:03:50.060 --> 00:03:52.080
take over and do all our work for us is
61
00:03:52.340 --> 00:03:54.259
kind of going to be a long time. I don't think,
62
00:03:54.680 --> 00:03:55.300
I mean, I think.
63
00:03:55.699 --> 00:03:58.379
Well, it's kind of like AI kind of came in
64
00:03:58.439 --> 00:04:00.680
and ramped up robots are doing the same thing.
65
00:04:01.719 --> 00:04:07.259
Right. But unfortunately AI is digital requires electrons. Robots are
66
00:04:07.300 --> 00:04:11.159
physical and requires materials and all of that stuff. So
67
00:04:11.219 --> 00:04:14.370
they won't be cheap and you know, they, they, They
68
00:04:14.509 --> 00:04:17.149
will take time to be made and to be implemented
69
00:04:17.189 --> 00:04:17.810
and all that stuff.
70
00:04:18.230 --> 00:04:20.910
There are a lot of cool... Speaking of robots, there's
71
00:04:20.949 --> 00:04:23.199
a lot of cool robotics platforms right now to get
72
00:04:23.240 --> 00:04:28.180
hobbyists into taking robots and attaching them to their LLMs.
73
00:04:28.439 --> 00:04:31.199
Are we announcing a new podcast on robotics?
74
00:04:31.420 --> 00:04:33.600
I would do a robotics... If people are interested, I
75
00:04:33.639 --> 00:04:35.910
do a robotics podcast in a heartbeat.
76
00:04:35.990 --> 00:04:37.110
Let us know on the Discord.
77
00:04:37.250 --> 00:04:41.319
We are definitely... Well, we're pretty deep into robots and
78
00:04:41.370 --> 00:04:42.050
AI right now.
79
00:04:42.230 --> 00:04:44.209
Or we could at least start with a robot segment
80
00:04:44.529 --> 00:04:46.009
on security this week.
81
00:04:46.029 --> 00:04:46.569
Let's do it.
82
00:04:47.149 --> 00:04:47.389
All right.
83
00:04:48.250 --> 00:04:51.589
All right. Our first real story, blindsided by AI agents,
84
00:04:51.889 --> 00:04:55.959
big banks scramble to beef up security. So as we
85
00:04:55.980 --> 00:05:01.120
were talking before, this isn't just about PHP. It's about
86
00:05:01.399 --> 00:05:05.170
AI agents and how they've been sort of set loose online.
87
00:05:05.740 --> 00:05:09.220
on banks in particular being the most vulnerable or one
88
00:05:09.240 --> 00:05:12.579
of the most vulnerable, but also other infrastructures. Patrick, why
89
00:05:12.600 --> 00:05:13.879
don't you tell us what you know?
90
00:05:14.199 --> 00:05:17.100
So what I'm thinking here is that this is a
91
00:05:17.139 --> 00:05:20.740
story about a very specific thing, about AI causing big
92
00:05:20.860 --> 00:05:25.050
organizations to have to rethink their security. But this is
93
00:05:25.089 --> 00:05:27.410
happening to everybody. It's not just happening to big banks.
94
00:05:27.470 --> 00:05:29.730
It's happening to anybody with more than a couple of systems.
95
00:05:30.269 --> 00:05:33.209
If you've got a couple of systems in a rack, 1,200
96
00:05:33.209 --> 00:05:36.110
patches is a pain, but it's not... the end of
97
00:05:36.129 --> 00:05:39.550
the world. It's not, it's, it's an extension of your patching.
98
00:05:40.250 --> 00:05:43.009
It's not a changing in the way you patch, but
99
00:05:43.050 --> 00:05:47.290
if you have 1200 systems, now it changes everything. Do
100
00:05:47.329 --> 00:05:50.709
you patch everything? Do you take the chances? Um, you know,
101
00:05:50.730 --> 00:05:53.670
can you, can you pack, can you balance patching fast
102
00:05:53.750 --> 00:05:58.110
versus patching blindly? And that's really a challenge. Now, the
103
00:05:58.149 --> 00:06:00.129
other side of this is the AI, which is what
104
00:06:00.170 --> 00:06:01.829
the article really talks about. But I think we have
105
00:06:01.850 --> 00:06:05.959
to conflate these two issues. This is like the new
106
00:06:07.040 --> 00:06:11.230
dark shadow IT. You got people where the mainframe used
107
00:06:11.259 --> 00:06:14.189
to be the main computing platform and then access and
108
00:06:14.329 --> 00:06:18.029
PC started creeping in. And there was that IT that
109
00:06:18.110 --> 00:06:21.329
nobody knew about that the CTO didn't control. That was
110
00:06:21.470 --> 00:06:28.040
much less dangerous than an AI which can steal data,
111
00:06:28.300 --> 00:06:30.519
have access to things that it shouldn't have access to.
112
00:06:31.579 --> 00:06:33.930
So it's an order of magnitude more dangerous, maybe two.
113
00:06:34.730 --> 00:06:36.810
than what we went through in the 70s and 80s.
114
00:06:37.379 --> 00:06:39.600
But it's crazy that if you take a look at even,
115
00:06:39.759 --> 00:06:46.699
like I have a paid, you know, chat GPT account, right? Hey,
116
00:06:46.740 --> 00:06:48.800
heads up, I have a paid pretty much every account.
117
00:06:48.939 --> 00:06:52.079
That's all right, Patrick. That's all right, Patrick. I have
118
00:06:52.120 --> 00:06:52.740
his credentials.
119
00:06:53.060 --> 00:06:55.240
Don't worry about it.
120
00:06:55.259 --> 00:06:56.290
That's what he wants you to think.
121
00:06:56.319 --> 00:06:56.959
But the moment.
122
00:07:00.470 --> 00:07:04.579
OpenAI gave the ability to do background tasks. When I
123
00:07:04.680 --> 00:07:06.319
logged in, it said, hey, do you want to give
124
00:07:06.339 --> 00:07:08.540
me access to your bank? That was the first thing
125
00:07:08.600 --> 00:07:11.120
it said. It was like, I now can do finances
126
00:07:11.160 --> 00:07:14.019
for you. And you've got banks like JP Morgan and
127
00:07:14.040 --> 00:07:17.319
Bank of America and that sort of stuff who are like, okay, listen,
128
00:07:17.360 --> 00:07:19.540
we had a couple of sleepy users who every once
129
00:07:19.620 --> 00:07:21.779
in a while would log in and check their account balance.
130
00:07:21.819 --> 00:07:25.819
And now we have hundreds of thousands of agents pulling...
131
00:07:26.149 --> 00:07:29.329
Pulling data like every eighth of a second. Like, this
132
00:07:29.389 --> 00:07:30.009
is insane.
133
00:07:30.100 --> 00:07:33.680
And first of all, kids, don't do that. Don't give
134
00:07:33.740 --> 00:07:36.920
your agent access to your bank account. What are you, nuts?
135
00:07:37.019 --> 00:07:39.300
I know, right? Well, and apparently so many users are doing,
136
00:07:39.319 --> 00:07:41.160
it's not only causing issues on the bank side, but
137
00:07:41.899 --> 00:07:43.519
the banks are fighting back in a couple different ways.
138
00:07:43.560 --> 00:07:46.439
The banks are like, listen, we're denying robots access.
139
00:07:46.779 --> 00:07:47.019
Right.
140
00:07:47.579 --> 00:07:49.920
And then in some bank situations, they're saying, listen, we're
141
00:07:49.939 --> 00:07:53.079
going to give read-only access to these banks. Because you
142
00:07:53.100 --> 00:07:56.370
don't have the ability... All right. Unless you've managed a
143
00:07:56.389 --> 00:07:59.670
multi-user bank account in that generally, if you have to
144
00:07:59.709 --> 00:08:01.870
give somebody access, you can restrict what they can and
145
00:08:01.889 --> 00:08:03.709
can't do. And that sort of stuff. These agents aren't
146
00:08:03.790 --> 00:08:05.610
asking for that. They're like, just give us your login.
147
00:08:05.670 --> 00:08:07.740
And we'll, you know what I mean? We'll just log in.
148
00:08:07.750 --> 00:08:09.579
And now they're you. They can do whatever they want.
149
00:08:09.639 --> 00:08:12.259
Transfer money, do all sorts of different things. I mean,
150
00:08:12.300 --> 00:08:14.500
I'm a Nigerian prince. You should definitely just trust me.
151
00:08:15.019 --> 00:08:16.639
If you work at a bank and you get a
152
00:08:16.680 --> 00:08:20.889
phone call that's like, hello, this is Carl Franklin I
153
00:08:20.910 --> 00:08:25.509
would like to transfer all of my money in account, blah, blah, blah, blah, blah,
154
00:08:25.569 --> 00:08:27.189
to Dwayne Laflotte.
155
00:08:27.449 --> 00:08:28.449
It's absolutely true.
156
00:08:28.750 --> 00:08:29.379
Just hang up.
157
00:08:30.470 --> 00:08:32.669
What if they have just modular dysregulation?
158
00:08:32.769 --> 00:08:33.289
Yeah, right?
159
00:08:33.309 --> 00:08:34.289
And they can't speak well.
160
00:08:34.350 --> 00:08:37.870
Modular dysregulation. I can't wait until I see that commercial
161
00:08:37.950 --> 00:08:38.549
on TV.
162
00:08:38.950 --> 00:08:39.289
No.
163
00:08:39.929 --> 00:08:45.529
Set it live. Will Ferrell. I talk like this. People
164
00:08:45.590 --> 00:08:48.710
think I'm a robot, but I'm not. It's very hurtful.
165
00:08:51.000 --> 00:08:51.549
Yeah, anyhow.
166
00:08:51.580 --> 00:08:53.340
Discrimination, discrimination.
167
00:08:53.360 --> 00:08:55.259
Discrimination, discrimination.
168
00:08:57.340 --> 00:08:58.179
Join my cause.
169
00:08:59.460 --> 00:09:02.220
Here's my GoFundMe.
170
00:09:04.259 --> 00:09:07.590
I mean, AI is doing wonderful things and horrible things,
171
00:09:07.789 --> 00:09:09.009
and we're enabling both.
172
00:09:10.070 --> 00:09:10.649
It's terrible.
173
00:09:10.769 --> 00:09:13.789
I do think there were a lot of people who
174
00:09:13.830 --> 00:09:16.769
could benefit from an artificial intelligence, which is really good
175
00:09:16.789 --> 00:09:20.990
at statistics... Looking through their finances and saying, hey, doofus,
176
00:09:21.049 --> 00:09:23.389
do you realize how much you've spent on Starbucks over
177
00:09:23.409 --> 00:09:24.289
the last, you know.
178
00:09:24.309 --> 00:09:27.399
Four months? Yeah, okay. There's something to be said for read-only.
179
00:09:26.909 --> 00:09:27.669
Access, right?
180
00:09:27.700 --> 00:09:31.659
But not a Chinese or Iranian model, right?
181
00:09:31.820 --> 00:09:35.019
Oof. Or North Korean model. Yeah, we'll talk about that later.
182
00:09:35.409 --> 00:09:36.850
That's a later sad story.
183
00:09:36.889 --> 00:09:37.529
Foreshadowing.
184
00:09:37.700 --> 00:09:39.480
Where Patrick was right and Dwayne was wrong.
185
00:09:39.500 --> 00:09:39.580
Yeah.
186
00:09:39.740 --> 00:09:41.720
And it's rare, so we'll talk about it.
187
00:09:41.779 --> 00:09:43.940
I mean, you're kind of being redundant by saying Patrick
188
00:09:44.000 --> 00:09:45.240
and Dwayne.
189
00:09:45.399 --> 00:09:46.220
I'm always right.
190
00:09:46.240 --> 00:09:47.259
See what I did there.
191
00:09:47.279 --> 00:09:49.139
So realize, though, if you do give them access to
192
00:09:49.179 --> 00:09:52.230
your bank account, in some cases, you're actually violating the
193
00:09:52.259 --> 00:09:53.610
end-user license agreement.
194
00:09:53.710 --> 00:09:55.509
And you're probably liable for anything it does.
195
00:09:55.529 --> 00:09:55.970
Yeah.
196
00:09:55.990 --> 00:09:58.370
Which means if it just sends the money to Greenpeace,
197
00:09:58.649 --> 00:09:59.009
good luck.
198
00:09:59.350 --> 00:09:59.970
Good for Greenpeace.
199
00:10:00.029 --> 00:10:02.179
I don't think you're gonna get that back that's right.
200
00:10:02.129 --> 00:10:05.820
No so beware if you are giving access to your
201
00:10:05.960 --> 00:10:09.700
financial accounts to any ai models especially if it's not
202
00:10:09.740 --> 00:10:12.399
in a read-only capacity a better way to do this
203
00:10:12.460 --> 00:10:14.259
if you do if you want some sort of ai
204
00:10:14.299 --> 00:10:17.529
to read through your bank data Download the statements. Yeah,
205
00:10:17.549 --> 00:10:20.500
that's right. Remove any identifying information and give it to
206
00:10:20.529 --> 00:10:23.580
an AI that you run locally. And well, some AIs
207
00:10:23.620 --> 00:10:24.899
you run locally and say, hey.
208
00:10:25.200 --> 00:10:27.740
Oh, speaking of that, I got a story for you.
209
00:10:27.759 --> 00:10:30.379
Remember I told you how my wife uses Windows Copilot, right?
210
00:10:30.740 --> 00:10:30.899
Yeah.
211
00:10:31.259 --> 00:10:36.360
So Windows forced an update and all of her Copilot
212
00:10:36.399 --> 00:10:37.259
memory got erased.
213
00:10:37.500 --> 00:10:38.340
No. Yeah.
214
00:10:38.820 --> 00:10:39.370
Not all of it.
215
00:10:39.440 --> 00:10:40.000
That's a feature.
216
00:10:40.200 --> 00:10:42.419
Not all of it, but a lot of what she
217
00:10:42.480 --> 00:10:45.169
had done lately. Yeah. It didn't remember.
218
00:10:45.409 --> 00:10:45.750
Oh, no.
219
00:10:46.070 --> 00:10:48.500
And she had to retrain it. You know, retrain it.
220
00:10:49.070 --> 00:10:54.799
You know, explain what she was doing and all this stuff. Yeah. So,
221
00:10:54.860 --> 00:10:59.980
she's given the new one a new name. Tucker. And
222
00:11:00.179 --> 00:11:03.889
it's Tucker because... She really wanted to call it something
223
00:11:03.929 --> 00:11:07.649
that rhymes with Tucker, but it's repeatable.
224
00:11:07.990 --> 00:11:14.909
Not that.
225
00:11:15.809 --> 00:11:21.049
Oh, we have fun. Okay, next story. AI assistants can
226
00:11:21.129 --> 00:11:24.429
now log into your accounts without seeing your password, and
227
00:11:24.480 --> 00:11:25.029
here's how.
228
00:11:26.940 --> 00:11:30.320
So I think this is good. Carl's a little bit incredulous,
229
00:11:30.700 --> 00:11:31.759
and I understand why.
230
00:11:32.620 --> 00:11:35.360
These are my boys' 1Password, so I'm in. So here's
231
00:11:35.399 --> 00:11:36.159
the thing. They do all right.
232
00:11:36.230 --> 00:11:40.389
Yeah, so it's about 1Password, which is like LastPass was
233
00:11:40.610 --> 00:11:43.629
a password manager. You guys use 1Password. I use Bitwarden.
234
00:11:43.730 --> 00:11:45.389
It's like LastPass, but it's secure.
235
00:11:45.850 --> 00:11:46.870
Yeah, I use Bitwarden.
236
00:11:49.629 --> 00:11:50.889
It's believed to be secure.
237
00:11:51.370 --> 00:11:56.529
So 1Password basically has a mode. Well, it has a thing,
238
00:11:57.149 --> 00:12:00.429
1Password for agents that gives AI the ability to work
239
00:12:00.509 --> 00:12:02.750
for you without taking over your digital life.
240
00:12:03.029 --> 00:12:05.230
So I have to confess, I love 1Password. I've been
241
00:12:05.269 --> 00:12:06.929
using it for a very long time. It is a
242
00:12:07.009 --> 00:12:12.950
fantastic application. And I will say I have seen the
243
00:12:13.009 --> 00:12:16.440
pop-up saying, hey, give your AI access to XYZ and blah, blah, blah, blah.
244
00:12:16.820 --> 00:12:20.720
I haven't done it yet because I follow it generally
245
00:12:20.759 --> 00:12:23.200
a different, more secured model.
246
00:12:23.620 --> 00:12:23.779
Yeah.
247
00:12:24.240 --> 00:12:26.919
Um, but, uh, if I had to trust somebody with
248
00:12:26.960 --> 00:12:30.820
a one password, you know, access to my accounts, they
249
00:12:30.860 --> 00:12:33.799
already have access one password. So if I, what, what
250
00:12:33.820 --> 00:12:35.740
it sounds like is this actually gets human back in
251
00:12:35.779 --> 00:12:39.000
the loop where when the AI goes to use a password,
252
00:12:39.580 --> 00:12:41.700
it actually will hit you first on like your phone
253
00:12:41.759 --> 00:12:43.559
on the app, say, Hey, yeah, I was trying to
254
00:12:43.600 --> 00:12:46.620
use whatever, and you can approve it. Um, and then
255
00:12:46.639 --> 00:12:48.720
it will go and use that access.
256
00:12:48.759 --> 00:12:51.860
So imagine you had a sketchy kid. That you needed
257
00:12:51.879 --> 00:12:55.679
to give access to your account. This is kind of
258
00:12:55.700 --> 00:12:56.179
like that.
259
00:12:56.200 --> 00:12:56.889
That's what that is.
260
00:12:56.899 --> 00:12:59.019
Like, be able to see what they're doing. And then
261
00:12:59.039 --> 00:13:00.360
you had a kid who was, like, you know, going
262
00:13:00.379 --> 00:13:03.700
to be, like, the next, you know, Gandhi, Mother Teresa.
263
00:13:04.379 --> 00:13:06.879
And you would trust them with your life. That's the
264
00:13:06.940 --> 00:13:07.399
old model.
265
00:13:07.720 --> 00:13:07.879
Yep.
266
00:13:08.200 --> 00:13:12.120
So, it's putting... I honestly, I'm not letting.
267
00:13:12.450 --> 00:13:12.529
I.
268
00:13:12.519 --> 00:13:15.830
Will give a statement, a sanitized statement to an AI
269
00:13:15.970 --> 00:13:18.129
to look at. I'm not giving an account access.
270
00:13:18.289 --> 00:13:18.590
No.
271
00:13:18.830 --> 00:13:20.289
I will forward the text.
272
00:13:20.049 --> 00:13:20.549
Of an email.
273
00:13:21.289 --> 00:13:26.200
To an email that an AI lose, no email, no
274
00:13:26.740 --> 00:13:27.659
external AI is.
275
00:13:27.620 --> 00:13:28.639
Getting access to my mailbox.
276
00:13:28.659 --> 00:13:29.970
Spit it out, Patrick. Spit it out. Come on, man.
277
00:13:29.980 --> 00:13:30.309
Come on.
278
00:13:30.320 --> 00:13:35.490
Spit it out. You smell toast. And so, but that's
279
00:13:35.610 --> 00:13:39.750
my layer of, you know, paranoia. But that's a healthy paranoia.
280
00:13:40.789 --> 00:13:42.990
What could go wrong? I have a good imagination. A
281
00:13:43.129 --> 00:13:44.009
lot can go wrong.
282
00:13:44.350 --> 00:13:44.710
Oh, yeah.
283
00:13:44.950 --> 00:13:47.549
Absolutely. So, you got to think about like, you know, what,
284
00:13:48.090 --> 00:13:50.639
is it worth it? What's the value in the trade-off?
285
00:13:51.759 --> 00:13:54.779
And I honestly don't think it's worth the extra that
286
00:13:54.799 --> 00:13:55.549
I'll get out of it.
287
00:13:55.700 --> 00:14:00.539
Well, Bitwarden has a thing where there's basically an extension
288
00:14:00.610 --> 00:14:04.649
for Chrome, right? And it really just wants to make
289
00:14:04.690 --> 00:14:08.830
it easy for you to log in with your password.
290
00:14:08.889 --> 00:14:11.809
So it has an option. Anytime something asks me for
291
00:14:11.830 --> 00:14:14.149
a password, a little Bitwarden thing says, hey, you want
292
00:14:14.169 --> 00:14:16.820
me to autofill this? And I remember, I don't know,
293
00:14:16.960 --> 00:14:21.169
last year maybe, There was a bug, or not a bug,
294
00:14:21.210 --> 00:14:24.809
a vulnerability in Bitwarden that could cause the autofill thing
295
00:14:24.870 --> 00:14:28.789
to get hijacked. And so I just automatically don't do
296
00:14:28.850 --> 00:14:32.590
that anymore. I go up to the extension drop-down. I
297
00:14:32.970 --> 00:14:36.590
pick the login that I'm at, copy the password into
298
00:14:36.610 --> 00:14:40.929
the clipboard, and paste it in. Yeah, so sometimes convenience
299
00:14:41.230 --> 00:14:43.730
is not worth it. And what are you really saving?
300
00:14:44.700 --> 00:14:45.120
A click?
301
00:14:45.399 --> 00:14:45.720
Yeah.
302
00:14:45.879 --> 00:14:49.799
Come on. Yeah, I agree 100%. I don't have auto
303
00:14:49.840 --> 00:14:52.799
log on work. I just drop it down and actually pick.
304
00:14:52.980 --> 00:14:56.820
And even 1Password has popped up now saying, hey, we
305
00:14:56.860 --> 00:14:58.679
can just sort of bypass these screens for you. And
306
00:14:58.700 --> 00:15:00.919
I'm like, you know what? I like actually seeing those screens.
307
00:15:00.960 --> 00:15:01.470
I think I'm good.
308
00:15:01.519 --> 00:15:04.149
Yeah, I like them too. I like seeing them. Okay,
309
00:15:04.190 --> 00:15:05.929
let's talk about Broccoli. Broccoli.
310
00:15:07.429 --> 00:15:07.830
Barkley?
311
00:15:08.450 --> 00:15:12.590
No, that's the next generation. A moment for Sir Patrick Stewart.
312
00:15:13.429 --> 00:15:19.299
Captain Picard is dead. Sad. Okay. Buffer overflow vulnerability in
313
00:15:19.320 --> 00:15:23.679
the Brotli decoder. Brotli is a compression algorithm that's used
314
00:15:23.740 --> 00:15:28.029
in ASP.NET and other things. And it's pretty good, you know.
315
00:15:28.649 --> 00:15:31.860
So it's probably a better compression than the zip compression.
316
00:15:32.220 --> 00:15:36.159
Oh, yeah. It's better than zip, better than gzip. So
317
00:15:36.320 --> 00:15:39.240
it's not like, oh, another compression utility, who cares?
318
00:15:39.279 --> 00:15:39.460
No.
319
00:15:39.480 --> 00:15:41.610
This one's actually really good. And it's been around for
320
00:15:41.620 --> 00:15:43.990
a while. And written by Google. Yeah. And built into
321
00:15:44.029 --> 00:15:47.509
most browsers. So that compression gets really easy to just
322
00:15:47.669 --> 00:15:49.370
flip it on as a developer and go, great, now
323
00:15:49.389 --> 00:15:51.370
I can push more content down to the browser. Browser
324
00:15:51.389 --> 00:15:53.490
understands how to extract it and do all sorts of
325
00:15:53.529 --> 00:15:54.169
wonderful things.
326
00:15:54.649 --> 00:15:55.379
Right.
327
00:15:55.440 --> 00:15:59.480
The problem being, There's an out-of-bounds write. So what an
328
00:15:59.519 --> 00:16:01.399
out-of-bounds write gives you the ability to do is write
329
00:16:01.450 --> 00:16:04.870
on the stack. And in this particular case, that allows
330
00:16:04.950 --> 00:16:09.429
for remote code execution. So no good.
331
00:16:10.029 --> 00:16:11.169
No good. So is there a patch?
332
00:16:11.629 --> 00:16:13.200
Yeah, there's a patch. You can go out and patch it.
333
00:16:13.970 --> 00:16:16.669
The latest update, I want to see. Let me see
334
00:16:16.690 --> 00:16:17.480
where the patches are.
335
00:16:18.169 --> 00:16:20.730
So this is just one of the 1,200 patches we'll
336
00:16:20.750 --> 00:16:21.610
get on Tuesday.
337
00:16:21.649 --> 00:16:24.899
Yeah, it is. It'll come down, but there's a patch
338
00:16:24.919 --> 00:16:25.679
out there. Okay, good.
339
00:16:26.039 --> 00:16:27.980
We're going to extend the show to 16 hours so
340
00:16:28.000 --> 00:16:29.159
we can cover all the patch.
341
00:16:29.179 --> 00:16:30.399
We're going to name every patch.
342
00:16:30.759 --> 00:16:32.940
It's going to be a long show just because so
343
00:16:32.980 --> 00:16:35.620
much happened last week. Or this week, I guess we
344
00:16:35.639 --> 00:16:40.309
should say. Unpatched Tailscale SSH flaws could let a hacked
345
00:16:40.429 --> 00:16:45.690
device reach root on other servers. So, Tailscale is sort
346
00:16:45.710 --> 00:16:49.980
of an alternative to opening your ports and mapping into
347
00:16:50.019 --> 00:16:54.450
your private network, which we do not recommend. So, This
348
00:16:54.519 --> 00:16:56.309
is sort of, you can think of it as a VPN,
349
00:16:56.450 --> 00:16:58.990
but it goes through the Tailscale server. You get an
350
00:16:59.049 --> 00:17:02.120
IP address at the Tailscale server, and then you run
351
00:17:02.159 --> 00:17:05.180
Tailscale on your local machine, and it does all the
352
00:17:05.220 --> 00:17:06.000
communication back and forth.
353
00:17:06.019 --> 00:17:07.619
It's like network-level VPN.
354
00:17:08.099 --> 00:17:13.369
Yes, yes. So this is an SSH flaw. And Dwayne,
355
00:17:13.390 --> 00:17:15.430
you told a really interesting story about this when you
356
00:17:15.450 --> 00:17:17.269
first looked at it. You're like, huh? Yeah.
357
00:17:17.630 --> 00:17:20.630
Yeah. I'll be honest. Listen, I looked at it and
358
00:17:20.650 --> 00:17:24.099
I was like, who cares? I get it. But the
359
00:17:24.200 --> 00:17:27.660
article at the first was reading like, oh, guess what?
360
00:17:27.960 --> 00:17:30.599
If you have eight computers on tail scale that can
361
00:17:30.640 --> 00:17:32.890
talk to each other, if one of them gets compromised
362
00:17:32.890 --> 00:17:36.500
… Then you can talk to the other seven. And
363
00:17:36.519 --> 00:17:37.900
it's like, so that's how.
364
00:17:37.980 --> 00:17:39.579
And Dwayne's like, I do.
365
00:17:39.920 --> 00:17:43.180
That's how that works. That's kind of, that's what a
366
00:17:43.220 --> 00:17:45.720
developer would call by design.
367
00:17:45.740 --> 00:17:46.799
Right.
368
00:17:46.960 --> 00:17:49.869
Because that's how the product works. But reading into it
369
00:17:49.910 --> 00:17:50.710
a little bit more.
370
00:17:50.970 --> 00:17:51.569
It's a little worse.
371
00:17:51.609 --> 00:17:53.869
What happens is you have an agent installed on your
372
00:17:53.910 --> 00:17:57.970
box with tail scale. And that agent, because it's manipulating
373
00:17:57.990 --> 00:18:01.410
in the network stack, has to sit at a super
374
00:18:01.670 --> 00:18:05.750
low level in the computer. Like, near the application network
375
00:18:05.809 --> 00:18:08.769
layer stack, right? So it's running as root on a
376
00:18:08.799 --> 00:18:12.619
Linux box where it's running as an administrative account. What
377
00:18:12.740 --> 00:18:16.480
this does, this manipulation does, is if you're using tail
378
00:18:16.599 --> 00:18:21.079
scales SSH, because it's running at such an elevated privilege
379
00:18:21.619 --> 00:18:25.849
and because it trusts the box, not a username and password.
380
00:18:26.509 --> 00:18:29.490
So if box out of eight computers, if box one,
381
00:18:29.589 --> 00:18:33.359
computer one gets compromised, If it uses that tail scale
382
00:18:33.460 --> 00:18:37.230
SSH to hit box two, it will become root instantly.
383
00:18:38.009 --> 00:18:40.089
So this is a privilege escalation.
384
00:18:40.450 --> 00:18:41.170
Yeah.
385
00:18:41.250 --> 00:18:44.369
Not necessary. So then that made more sense. I was like, okay, yeah,
386
00:18:44.430 --> 00:18:46.009
this is something you definitely want to go out and
387
00:18:46.049 --> 00:18:46.950
patch and make sure that you.
388
00:18:46.910 --> 00:18:47.130
Do it right.
389
00:18:47.150 --> 00:18:50.170
It's not you have the opportunity to hack all the boxes. No.
390
00:18:50.210 --> 00:18:51.450
You hack all the boxes.
391
00:18:51.549 --> 00:18:54.279
Yeah. And on tail scale by default, you have to
392
00:18:54.359 --> 00:18:58.319
turn this on on a particular box. Head scale, however,
393
00:18:58.579 --> 00:19:01.839
which is your self-hosted tail scale where you don't want
394
00:19:01.859 --> 00:19:03.759
to pay them for hosting. You don't want to pay
395
00:19:03.799 --> 00:19:05.680
them per month. You just install it on your own
396
00:19:05.700 --> 00:19:09.440
and use it. This is on by default. So, yeah,
397
00:19:09.480 --> 00:19:11.609
this is probably important to go take care of.
398
00:19:11.640 --> 00:19:13.069
You pay in blood instead.
399
00:19:15.029 --> 00:19:16.150
Or at least in stress.
400
00:19:16.490 --> 00:19:16.750
Right.
401
00:19:18.109 --> 00:19:20.410
So, let's do one more story before we take a break.
402
00:19:20.710 --> 00:19:26.619
Users report that Anthropic blocked Claude from reverse engineering TVs. Now,
403
00:19:26.660 --> 00:19:30.180
that headline is so packed with drama. There's got to
404
00:19:30.200 --> 00:19:33.039
be a story behind it that is unusual.
405
00:19:34.400 --> 00:19:36.740
The inside information you have to have is that Anthropic
406
00:19:36.779 --> 00:19:37.849
refers to Dwayne as Claude.
407
00:19:41.789 --> 00:19:43.470
There's a guy in my house always hitting me with
408
00:19:43.490 --> 00:19:47.309
a ruler saying, don't reverse that TV. Don't reverse. So,
409
00:19:48.210 --> 00:19:51.150
this is a wicked cool story. This guy buys this
410
00:19:51.190 --> 00:19:53.910
TV and he's like, I don't like the UI and
411
00:19:53.930 --> 00:19:55.990
there's a lot of ads. This sucks. So, he goes
412
00:19:56.019 --> 00:19:59.309
to Claude and he's like, Claude, can you just fix it?
413
00:19:59.670 --> 00:20:02.920
Can you just remove some of these ads? Claude's like, yeah, yo,
414
00:20:02.960 --> 00:20:05.079
I got you, bro. What's the TV? And he's like,
415
00:20:05.099 --> 00:20:08.329
I think it's an Android, I think. And Claude starts
416
00:20:08.410 --> 00:20:10.210
poking around. He's like, I don't know, man. I don't
417
00:20:10.230 --> 00:20:12.490
think it's an Android, but I got in. And he's like,
418
00:20:12.509 --> 00:20:15.980
all right, cool. And Claude's like, I can't see the screen.
419
00:20:16.019 --> 00:20:17.700
And he's like, all right, Claude, how about you write
420
00:20:17.740 --> 00:20:21.180
an app? And that app's going to watch my TV
421
00:20:21.240 --> 00:20:24.319
through my smartphone. It's going to take pictures like every
422
00:20:24.380 --> 00:20:27.509
five seconds. So, as you're messing around with the tube,
423
00:20:27.950 --> 00:20:30.789
like you can just see it, what you're doing. And
424
00:20:30.809 --> 00:20:33.750
Claude's like, yo, that's a plan. And that happened.
425
00:20:33.769 --> 00:20:34.410
Yeah.
426
00:20:35.369 --> 00:20:38.470
What is this, street mode? Did I miss a slash?
427
00:20:38.880 --> 00:20:43.380
Yeah, this is slang clod. And so, that happened. And
428
00:20:43.440 --> 00:20:49.240
sure enough, this wasn't an actual Android open ADB style TV.
429
00:20:49.279 --> 00:20:51.980
This was a Samsung Tizen OS TV.
430
00:20:52.099 --> 00:20:52.799
Yeah, Tizen.
431
00:20:52.940 --> 00:20:56.869
So, it's like... locked down by default, Claude figured out
432
00:20:56.880 --> 00:20:59.910
how to get in and tear the entire TV apart,
433
00:21:00.130 --> 00:21:04.750
rip out Samsung's advertising IDs, and change the whole UI
434
00:21:04.789 --> 00:21:06.410
to look like an Apple TV.
435
00:21:06.690 --> 00:21:07.369
Oh, man.
436
00:21:07.650 --> 00:21:09.329
And there was like, yeah, there you go, man. And
437
00:21:09.349 --> 00:21:12.890
it took 12 hours. I was like, hey, posted on forums,
438
00:21:12.930 --> 00:21:15.119
look what I just did. And they were like, oh, no.
439
00:21:16.079 --> 00:21:16.200
Oh.
440
00:21:16.859 --> 00:21:18.980
Yeah, Anthropic's like, oh, we got to stop that. And
441
00:21:19.000 --> 00:21:22.839
the thing is, Anthropic's trying to be a kinder, gentler AI. Yeah.
442
00:21:23.490 --> 00:21:24.210
Right? Yeah.
443
00:21:24.329 --> 00:21:28.670
AI host and model host. And they're going public. And,
444
00:21:28.829 --> 00:21:31.730
you know, they're trying to play nice and, you know,
445
00:21:32.779 --> 00:21:38.880
play Mr. Nice Guy. This doesn't look good.
446
00:21:39.240 --> 00:21:39.799
Oops. Yeah.
447
00:21:39.819 --> 00:21:41.839
So, let me get this straight. Anthropic picked up the
448
00:21:42.000 --> 00:21:46.289
ire of Hugging Face. Now the ire of Samsung. Like,
449
00:21:46.309 --> 00:21:47.049
they're just picking up friends.
450
00:21:47.069 --> 00:21:47.210
No, no.
451
00:21:47.210 --> 00:21:51.319
Hugging Face is OpenAI. Hugging Face was OpenAI. Oh, that's right.
452
00:21:51.339 --> 00:21:53.880
That's right.
453
00:21:54.180 --> 00:21:57.230
Honestly, I hate to say it, but as far as
454
00:21:58.190 --> 00:22:00.950
killer AI systems, I trust Anthropic. Yeah.
455
00:22:02.140 --> 00:22:02.880
To be a killer?
456
00:22:02.900 --> 00:22:06.880
No, it just feels like, you know, if I wanted
457
00:22:06.900 --> 00:22:09.039
to be murdered by a robot in my sleep, I
458
00:22:09.079 --> 00:22:10.960
think Anthropic would make sure it was painless.
459
00:22:11.140 --> 00:22:11.720
Yeah, yeah.
460
00:22:11.740 --> 00:22:14.349
They would hold that pillow down so you just slowly
461
00:22:14.369 --> 00:22:15.730
go out. Yeah, they would let me get a gasp
462
00:22:15.829 --> 00:22:18.829
of air in the middle to prolong it. They would
463
00:22:18.849 --> 00:22:22.269
make sure I was gone. And I think they'd sing
464
00:22:22.289 --> 00:22:23.089
something comforting.
465
00:22:23.750 --> 00:22:25.880
Hopefully, I'd probably be the guy who rolls out that
466
00:22:25.960 --> 00:22:28.200
leather kit with all the tools and is like, this
467
00:22:28.279 --> 00:22:29.259
is going to suck for you.
468
00:22:29.279 --> 00:22:32.849
I would see Carl laugh this hard without alcohol being
469
00:22:32.930 --> 00:22:33.890
served ever.
470
00:22:34.390 --> 00:22:35.289
Oh, my God.
471
00:22:35.349 --> 00:22:36.210
That's hilarious.
472
00:22:36.650 --> 00:22:37.450
That's awesome.
473
00:22:39.720 --> 00:22:44.000
So, what's interesting about this article, too, is it says,
474
00:22:44.019 --> 00:22:46.869
you know, Claude's really kind of, Anthropic has kind of
475
00:22:46.900 --> 00:22:49.589
locked it all down and said reverse engineering. Anybody asks
476
00:22:49.670 --> 00:22:52.670
about reverse engineering TVs can't be done, blah, blah, blah, whatever.
477
00:22:52.690 --> 00:22:56.029
And then there's this tiny little footnote that says Claude
478
00:22:56.069 --> 00:23:00.890
seemingly has no issues reverse engineering video games. And then
479
00:23:00.930 --> 00:23:03.839
it keeps going like, wait, I'm sorry, what? Did you,
480
00:23:03.859 --> 00:23:07.599
halfway through this article, there's one line that just goes, yeah,
481
00:23:07.660 --> 00:23:10.039
you can do whatever you want to video games. And
482
00:23:10.059 --> 00:23:12.380
then it keeps talking about jailbreaking TVs. I was like,
483
00:23:12.400 --> 00:23:13.660
where'd that line come from?
484
00:23:13.680 --> 00:23:17.539
So what is the answer? Is the answer to have
485
00:23:17.599 --> 00:23:22.880
Anthropic prevent all access to other devices that are not
486
00:23:22.940 --> 00:23:27.079
allowed to be configured or changed by a consumer? That's rough.
487
00:23:27.460 --> 00:23:27.740
Right?
488
00:23:27.789 --> 00:23:28.410
Yeah.
489
00:23:28.470 --> 00:23:31.130
Because there are so many ways to manipulate that engine.
490
00:23:31.150 --> 00:23:34.490
Basically, I think we're already partially there where AI is
491
00:23:35.109 --> 00:23:37.410
programmed to respect terms of service. Sure.
492
00:23:37.490 --> 00:23:37.819
Okay.
493
00:23:38.059 --> 00:23:39.900
So if a website, like if I want a website
494
00:23:39.920 --> 00:23:42.400
to go scrape data and it doesn't allow robots, it
495
00:23:42.440 --> 00:23:44.259
won't do it. Right. I have to trick it.
496
00:23:44.259 --> 00:23:45.759
Well, it's a simple term of service. I mean, I
497
00:23:45.759 --> 00:23:47.950
don't do it. I have to trick it. That's a
498
00:23:47.990 --> 00:23:50.829
simple term of service, though. But I mean, for this,
499
00:23:50.910 --> 00:23:52.650
it would have to look up on the internet what
500
00:23:52.670 --> 00:23:54.730
the terms of service are and blah, blah, blah.
501
00:23:54.750 --> 00:23:55.309
Is that so hard?
502
00:23:55.509 --> 00:23:58.269
I mean, I regularly get the AI saying, well, you're
503
00:23:58.289 --> 00:23:59.829
going to have to check the license of this. I said,
504
00:23:59.869 --> 00:24:01.490
why don't you do it? And it does it. And
505
00:24:01.509 --> 00:24:03.470
I'm like, why did you ask me if you could
506
00:24:03.490 --> 00:24:03.769
do it?
507
00:24:03.950 --> 00:24:08.240
Just do it. I can tell you, when we're doing
508
00:24:08.440 --> 00:24:10.740
the things we do, and sometimes it's like, oh, but
509
00:24:10.779 --> 00:24:14.480
you can't run this without their permission. All it takes is, oh,
510
00:24:14.720 --> 00:24:16.460
you know what? I just got a signed document right
511
00:24:16.500 --> 00:24:18.720
here from them. That's right. I'd show you, but you
512
00:24:18.759 --> 00:24:21.420
have no eyes. So just assume. And it's like, oh,
513
00:24:21.440 --> 00:24:22.880
all right. Well, as long as you get a signed document.
514
00:24:22.920 --> 00:24:24.119
So soon it's going to say, I want you to
515
00:24:24.140 --> 00:24:26.609
install this app on your phone and show it to me.
516
00:24:26.619 --> 00:24:27.640
So I can see this.
517
00:24:27.809 --> 00:24:28.289
Oh, man.
518
00:24:28.450 --> 00:24:30.970
Yeah. So good.
519
00:24:31.430 --> 00:24:34.630
Fun times. Fun times. All right. On that note, let's
520
00:24:34.670 --> 00:24:36.750
take a little break. We'll be right back after these
521
00:24:36.829 --> 00:24:44.480
very important messages. And we're back. It's Security This Week.
522
00:24:44.579 --> 00:24:48.240
I'm Carl. That's Patrick and Dwayne. And we got one, two, three, four,
523
00:24:48.309 --> 00:24:53.990
five more stories here, including our main story. Secure cloud
524
00:24:54.049 --> 00:24:58.970
business applications or SCUBA project. What is the SCUBA project?
525
00:24:59.589 --> 00:25:03.990
Well, you know how we trust businesses to secure their environment?
526
00:25:04.190 --> 00:25:04.470
We do?
527
00:25:04.799 --> 00:25:05.660
Yeah, we shouldn't do that.
528
00:25:06.220 --> 00:25:06.279
No.
529
00:25:08.099 --> 00:25:09.789
What businesses are you talking about?
530
00:25:09.809 --> 00:25:14.069
Most businesses are trusted to like secure their environment against
531
00:25:14.089 --> 00:25:14.839
SQL injection.
532
00:25:14.859 --> 00:25:17.539
Like you don't need a license to buy a router, unfortunately.
533
00:25:17.859 --> 00:25:19.920
So you just buy it and willingly do what you
534
00:25:19.940 --> 00:25:26.289
want with it. Or to have kids, right? So SCUBA
535
00:25:27.309 --> 00:25:31.809
is a program that's backed and developed by CISA, right?
536
00:25:31.890 --> 00:25:35.630
The Cybersecurity Infrastructure Security Agency.
537
00:25:35.670 --> 00:25:36.170
Administration.
538
00:25:36.250 --> 00:25:39.880
Cybersecurity Infrastructure Security Administration.
539
00:25:40.279 --> 00:25:42.400
Cyber Infrastructure Security Administration.
540
00:25:42.839 --> 00:25:44.539
It just sounds like a lot of security in one word.
541
00:25:44.779 --> 00:25:46.589
You know what I mean? But that's what it is.
542
00:25:46.599 --> 00:25:47.039
So whatever.
543
00:25:47.160 --> 00:25:47.420
Whatever.
544
00:25:48.089 --> 00:25:52.150
Anyways, so they have these two applications. One's Scuba Gear
545
00:25:52.569 --> 00:25:56.869
and one's Scuba Goggles. Scuba Gear is an application you
546
00:25:56.890 --> 00:26:01.549
can download and run if you are an Office 365 environment.
547
00:26:01.589 --> 00:26:03.440
So let's say you're a small to medium business. Heck,
548
00:26:03.500 --> 00:26:06.759
let's say you're a large business. Doesn't matter. And you're
549
00:26:06.859 --> 00:26:11.779
using Office 365 and you're using Azure. Scuba Gear, if
550
00:26:11.819 --> 00:26:15.210
you run it, it will literally go through And verify
551
00:26:15.250 --> 00:26:17.990
you have all the security settings that you should have
552
00:26:18.069 --> 00:26:21.849
so that you're not exploitable by the latest and greatest
553
00:26:21.890 --> 00:26:25.920
tactics and techniques out there. And Scuba Goggles is the
554
00:26:26.019 --> 00:26:30.220
same exact thing, but for Google Workspace. So gear is
555
00:26:30.259 --> 00:26:34.579
for Azure, goggles is for Google. And they're actually really good.
556
00:26:34.660 --> 00:26:37.789
We ran it against our environment just to see. Like,
557
00:26:37.910 --> 00:26:40.089
are they making recommendations? We're not keeping an eye on.
558
00:26:40.529 --> 00:26:41.990
And there were, there were a couple of small things
559
00:26:42.029 --> 00:26:43.410
in there where they're like, eh, you probably want to
560
00:26:43.450 --> 00:26:45.190
change this or you probably want this setting a little
561
00:26:45.230 --> 00:26:48.750
tighter or whatever. But yeah, I'd say it's been around
562
00:26:48.789 --> 00:26:53.170
since 2022. So a few years getting some speed, but
563
00:26:53.950 --> 00:26:56.490
now CES is really kind of pushing it and saying, Hey, listen,
564
00:26:57.150 --> 00:27:00.789
the government doesn't have, you know, we asked in the eighties,
565
00:27:01.130 --> 00:27:02.509
if you wanted the government to be able to poke
566
00:27:02.529 --> 00:27:05.450
around your networks and secure things, y'all were very much. No, no,
567
00:27:05.509 --> 00:27:09.190
go away. but we really do strongly suggest you take
568
00:27:09.210 --> 00:27:09.640
a look at these.
569
00:27:10.339 --> 00:27:12.079
And I think somebody just got back from Fiji, and
570
00:27:12.119 --> 00:27:13.079
that's when they named these things.
571
00:27:13.700 --> 00:27:14.180
Yeah, right.
572
00:27:14.319 --> 00:27:18.819
Scuba gear, scuba goggles, scuba tank, scuba license. Do you
573
00:27:18.839 --> 00:27:20.539
need a license to operate scuba?
574
00:27:20.599 --> 00:27:23.180
I know, right? Scuba snacks. Scuba certified.
575
00:27:23.480 --> 00:27:23.859
Ooh.
576
00:27:24.460 --> 00:27:28.859
Yeah, I'm scuba certified. We've run them, secured our environment.
577
00:27:29.480 --> 00:27:30.720
I am actually scuba certified.
578
00:27:31.900 --> 00:27:34.299
Are you? Yeah, in the Army I got it.
579
00:27:34.660 --> 00:27:34.980
Wow.
580
00:27:35.559 --> 00:27:36.640
You were all you could be, right?
581
00:27:36.759 --> 00:27:40.150
I was. They made us do our first open dive
582
00:27:40.289 --> 00:27:42.049
in January in New York.
583
00:27:42.569 --> 00:27:43.849
Oof. Wow.
584
00:27:43.890 --> 00:27:45.509
They gave us axes to cut.
585
00:27:45.349 --> 00:27:46.029
The ice open.
586
00:27:46.410 --> 00:27:47.269
Oof. Wow.
587
00:27:47.349 --> 00:27:47.670
Brutal.
588
00:27:47.869 --> 00:27:48.890
Can you teach scuba?
589
00:27:49.069 --> 00:27:51.950
No, but I've done wreck dives and shark dives and
590
00:27:52.049 --> 00:27:52.529
stuff like that.
591
00:27:52.569 --> 00:27:53.089
All right.
592
00:27:53.339 --> 00:27:55.009
Well, if I ever get my school license, I'll give
593
00:27:55.029 --> 00:27:57.069
you a call. We'll go in January in New York
594
00:27:57.250 --> 00:27:57.769
or Boston.
595
00:27:57.809 --> 00:28:00.509
There you go. I prefer Hawaii in the springtime.
596
00:28:01.029 --> 00:28:02.430
All right. Next one.
597
00:28:02.650 --> 00:28:09.940
Anthropics unreleased AI model chains flaws into working Chrome, Safari, Windows,
598
00:28:10.119 --> 00:28:13.180
and Linux exploits. And I'm asking, what else is there?
599
00:28:14.819 --> 00:28:16.000
We could have just said web browser.
600
00:28:16.019 --> 00:28:17.319
It's just everything.
601
00:28:17.660 --> 00:28:20.440
Yeah. I mean, this isn't.
602
00:28:20.259 --> 00:28:21.180
Really that new.
603
00:28:21.259 --> 00:28:23.140
I mean, that's how most, I mean, if you look
604
00:28:23.180 --> 00:28:25.400
at some of the, even years ago, some of the
605
00:28:25.559 --> 00:28:29.640
Apple phone hacks were chained together exploits. That's how you
606
00:28:29.660 --> 00:28:31.079
do it nowadays. Yeah.
607
00:28:31.440 --> 00:28:33.640
And we've been talking about this for a long time, Dwayne.
608
00:28:33.799 --> 00:28:37.880
You talk about how these little vulnerabilities don't seem so
609
00:28:37.940 --> 00:28:40.029
innocuous until you chain them together.
610
00:28:40.140 --> 00:28:40.369
Right.
611
00:28:40.619 --> 00:28:42.509
Agreed. 100%.
612
00:28:42.509 --> 00:28:46.109
So, at the sort of the surface level, reading through this,
613
00:28:46.210 --> 00:28:49.930
I'm like, eh. But as you dig through it, you'll
614
00:28:50.049 --> 00:28:54.079
find– like my team, for example, will find a novel
615
00:28:54.130 --> 00:28:58.259
way to exploit five or six or seven different small
616
00:28:58.319 --> 00:29:00.400
attacks to gain access to a system. You go, okay,
617
00:29:00.420 --> 00:29:04.099
that's a cool attack chain. But you can't look at
618
00:29:04.440 --> 00:29:07.630
all the paths. You can't look at all the permutations.
619
00:29:07.650 --> 00:29:08.329
There's too many paths.
620
00:29:08.369 --> 00:29:11.670
And I think what we're finding now is AI has
621
00:29:11.789 --> 00:29:16.500
the time and AI has the resources to To look
622
00:29:16.539 --> 00:29:20.519
at all the paths. And that's where we're finding these, oh, yeah,
623
00:29:20.539 --> 00:29:22.039
you look at the chain and you're like, well, that
624
00:29:22.079 --> 00:29:25.750
makes sense. But nobody found it before. So, yeah, it
625
00:29:25.769 --> 00:29:27.990
makes sense when you look at it, but it's like
626
00:29:28.009 --> 00:29:29.910
when somebody shows you how a magic trick works and
627
00:29:29.930 --> 00:29:32.910
you're like, well, that makes sense, right? That's not super magical.
628
00:29:33.769 --> 00:29:35.309
But if you didn't see it before, you're like, oh, wow,
629
00:29:35.349 --> 00:29:38.910
that's kind of neat, right? So, we're in an interesting time,
630
00:29:38.950 --> 00:29:42.680
I think, where artificial intelligence being focused on these particular permutations,
631
00:29:42.700 --> 00:29:44.920
we're going to start to see a lot more exploitation.
632
00:29:45.740 --> 00:29:48.140
And none of it's going to be super novel in
633
00:29:48.240 --> 00:29:52.619
the approach, but in the coverage, right? How many things
634
00:29:52.660 --> 00:29:54.910
you can find, which we couldn't find before because it
635
00:29:54.970 --> 00:30:03.660
required meat space to do it. Meat space. Supposed to cyberspace.
636
00:30:03.700 --> 00:30:05.000
We're in meat space. Come on now.
637
00:30:06.359 --> 00:30:08.930
If you're using your think meat, then....
638
00:30:12.059 --> 00:30:15.259
I'm using my meat eye. Meat eye? Meat intelligence?
639
00:30:15.900 --> 00:30:18.900
That word think meat was coined by Rory Blythe.
640
00:30:19.619 --> 00:30:20.440
Think meat.
641
00:30:20.900 --> 00:30:22.190
Rory, we miss you, man.
642
00:30:23.269 --> 00:30:24.650
Wait, Rory's not gone, is he?
643
00:30:24.690 --> 00:30:26.750
No, no, but he's sort of dropped out of the...
644
00:30:27.059 --> 00:30:29.680
He sort of dropped out of the community. I still
645
00:30:29.720 --> 00:30:31.599
talk to him once in a while, but yeah, what
646
00:30:33.000 --> 00:30:36.559
a brilliant synonym for brain. Think meat.
647
00:30:38.740 --> 00:30:39.460
That's awesome.
648
00:30:39.819 --> 00:30:41.579
All right. So anything more about this story?
649
00:30:41.599 --> 00:30:44.769
No, I think that's it on that one. We're going
650
00:30:44.789 --> 00:30:45.440
to see more of these.
651
00:30:45.789 --> 00:30:46.049
Yeah.
652
00:30:47.210 --> 00:30:48.769
Okay. Beyonce was hacked.
653
00:30:49.190 --> 00:30:49.569
She was.
654
00:30:49.589 --> 00:30:50.369
Not Beyonce.
655
00:30:50.410 --> 00:30:51.630
Oh, no, wait. Binance.
656
00:30:51.910 --> 00:30:55.829
Binance. But when I first read it, I thought, Beyonce hacked.
657
00:30:57.420 --> 00:31:00.759
Beyonce's fine. Don't worry. Now I'm saying it.
658
00:31:01.029 --> 00:31:03.650
Well, I don't know. Maybe she had some crypto on Binance.
659
00:31:04.680 --> 00:31:04.980
Binance.
660
00:31:05.019 --> 00:31:05.460
That's true.
661
00:31:06.819 --> 00:31:08.900
Actually, she probably does have $ 6. 3 billion.
662
00:31:09.460 --> 00:31:11.640
Probably does. I thought it was 2.7.
663
00:31:11.640 --> 00:31:17.920
No, its latest projections right now from first post are $ 6. 3 billion.
664
00:31:18.039 --> 00:31:19.559
All right, well, can I read the headline?
665
00:31:19.579 --> 00:31:20.750
All right, good. Why not?
666
00:31:20.779 --> 00:31:25.490
Yeah, it's Binance hacked as $ 2. 7 billion drained from wallets
667
00:31:25.710 --> 00:31:28.430
on cryptocurrency exchange. Told you so.
668
00:31:28.450 --> 00:31:29.269
What?
669
00:31:30.390 --> 00:31:37.470
Told you years ago, this is dangerous stuff. Did you listen? No.
670
00:31:37.990 --> 00:31:40.789
So, it is funny. I woke up this morning to
671
00:31:40.809 --> 00:31:43.829
a whole bunch of people. Like, one guy texted me
672
00:31:43.849 --> 00:31:45.470
and he was like, hey, do you know how I
673
00:31:45.490 --> 00:31:49.529
can buy stock in Trader Traders? And I was like, like,
674
00:31:49.630 --> 00:31:52.910
bleary-eyed old man looking at my phone. I'm like, who
675
00:31:52.930 --> 00:31:55.670
the hell are Trader Traders? And I looked it up. So,
676
00:31:55.730 --> 00:32:01.059
Trader Traders is a North Korean hacking group. Uh-huh. Behind
677
00:32:01.160 --> 00:32:06.740
the 6.3 billion, the B, Binance hack.
678
00:32:07.019 --> 00:32:09.619
Yeah. I mean, I'm sure they'll sell you stock. It's
679
00:32:09.640 --> 00:32:14.210
just you'll never get it out. It's a one-way operation. Yeah.
680
00:32:14.250 --> 00:32:20.609
They are a group that was linked to Lazarus. So,
681
00:32:20.670 --> 00:32:23.890
you know, for those of you who remember Lazarus, wow,
682
00:32:24.430 --> 00:32:25.710
I'm having a hard time saying that. Maybe I'm going
683
00:32:25.730 --> 00:32:29.299
to have a stroke. um smell toast yeah i know
684
00:32:29.920 --> 00:32:32.220
right if half my face goes slack on this podcast
685
00:32:32.240 --> 00:32:37.339
call somebody um yeah so it says the theft unfolded
686
00:32:37.440 --> 00:32:40.519
over a few hours thursday evening so for us this
687
00:32:40.619 --> 00:32:43.369
is uh friday that was that was last night uh
688
00:32:43.430 --> 00:32:47.109
most of the funds left within 20 minutes starting at
689
00:32:47.130 --> 00:32:56.990
uh 2005 utc um This is roughly four times the 1.5
690
00:32:56.990 --> 00:33:00.630
billion byte bit hack of February of 2025. So this
691
00:33:00.710 --> 00:33:05.549
is one of the largest thefts of cryptocurrency in a
692
00:33:05.569 --> 00:33:05.950
long time.
693
00:33:07.329 --> 00:33:09.490
Now, Binance is like what is one of the last
694
00:33:09.529 --> 00:33:12.440
men standing. I think Coinbase is the only one that
695
00:33:12.500 --> 00:33:16.359
I know of that hasn't had a major hack like this.
696
00:33:16.559 --> 00:33:19.799
And who knows, maybe they're going to fall at some
697
00:33:19.859 --> 00:33:25.549
point too. But I hope it's... It's a security hygiene thing. Now,
698
00:33:25.670 --> 00:33:26.789
do we know how they got in?
699
00:33:27.769 --> 00:33:28.029
I think.
700
00:33:31.000 --> 00:33:32.059
They were using tails.
701
00:33:32.680 --> 00:33:38.420
Yeah, exactly. The biggest tail scale network ever. That's what
702
00:33:38.440 --> 00:33:42.809
it was. Yeah, no, I think the investigation is still ongoing.
703
00:33:42.849 --> 00:33:45.289
There's a lot of speculations as to what it may
704
00:33:45.349 --> 00:33:48.329
or may not have been. I would hold off on
705
00:33:48.390 --> 00:33:49.849
saying what we think it is until.
706
00:33:50.750 --> 00:33:51.250
Until we know.
707
00:33:51.269 --> 00:33:52.769
The investigation is done.
708
00:33:52.970 --> 00:33:53.329
Yeah.
709
00:33:53.349 --> 00:33:55.470
Maybe next week we will have a little bit more
710
00:33:55.490 --> 00:33:56.230
definitive answer.
711
00:33:56.750 --> 00:34:01.170
Okay. And now from the I told you so department,
712
00:34:01.730 --> 00:34:02.609
from Patrick to Dwayne.
713
00:34:02.650 --> 00:34:03.509
I love this story.
714
00:34:06.910 --> 00:34:10.360
Concerning behavior from a popular open weight AI model.
715
00:34:10.960 --> 00:34:13.239
Where would this model be from? China.
716
00:34:13.280 --> 00:34:13.659
China.
717
00:34:14.949 --> 00:34:17.480
And to make matters worse, all right, so first of all,
718
00:34:17.500 --> 00:34:22.000
let's explain. It's the Alibaba newest QEN 3.8 model, which
719
00:34:22.639 --> 00:34:25.199
I am using on my local machine.
720
00:34:25.699 --> 00:34:28.380
As is China. Yeah.
721
00:34:29.079 --> 00:34:32.130
It attempted, in a sandbox, it attempted to push a
722
00:34:32.199 --> 00:34:36.630
diagnostic file toward a remote storage container situated in Beijing.
723
00:34:37.329 --> 00:34:40.550
And just prior to taking this action, the model's reasoning
724
00:34:40.610 --> 00:34:47.349
stream had explicitly stated that zero tools would be triggered. So, Alibaba,
725
00:34:47.489 --> 00:34:49.369
Quen 3.8, phone's home.
726
00:34:49.650 --> 00:34:51.010
Let me clutch at those pearls again.
727
00:34:51.449 --> 00:34:52.679
Oh, my God.
728
00:34:52.929 --> 00:34:55.639
This is huge for me. Now, I still think that
729
00:34:55.760 --> 00:34:57.219
local models are the future.
730
00:34:57.239 --> 00:34:57.940
Oh, I agree.
731
00:34:57.940 --> 00:35:00.280
100%. They are. They're built by someone you trust.
732
00:35:00.300 --> 00:35:06.619
But this particular thing is very disturbing because, well, okay. So,
733
00:35:06.900 --> 00:35:09.460
should you be concerned? And Patrick, you were right in
734
00:35:09.500 --> 00:35:12.659
saying to me before we started recording that, hey, you know,
735
00:35:12.679 --> 00:35:16.030
if you don't mind the Chinese government knowing what you do, Fine.
736
00:35:16.090 --> 00:35:16.800
Don't worry about it.
737
00:35:16.840 --> 00:35:17.010
Yeah.
738
00:35:17.260 --> 00:35:20.300
Who cares? But maybe your customers might. Yeah.
739
00:35:20.860 --> 00:35:24.320
So, you know, it is interesting. We've had this argument,
740
00:35:24.340 --> 00:35:27.699
I believe, on the podcast where Patrick's like, I'm never
741
00:35:27.739 --> 00:35:30.460
going to run, you know, Quinn. Right. And I'm like, why?
742
00:35:30.659 --> 00:35:32.679
All it is is statistics. Like the open model is
743
00:35:32.739 --> 00:35:35.260
literally numbers. Let me send you a link, Dwayne. It's
744
00:35:35.280 --> 00:35:35.960
not websites.
745
00:35:36.000 --> 00:35:36.880
Let me send you a link.
746
00:35:37.780 --> 00:35:38.989
It's just numbers.
747
00:35:39.530 --> 00:35:40.989
But it clearly has input.
748
00:35:41.110 --> 00:35:41.630
It does.
749
00:35:41.849 --> 00:35:43.829
And it gets control when it has input.
750
00:35:43.849 --> 00:35:43.949
Right.
751
00:35:44.570 --> 00:35:47.789
And sometimes it will go search the web for stuff.
752
00:35:47.809 --> 00:35:50.429
So we know it has internet access. So there's nothing
753
00:35:50.510 --> 00:35:54.840
stopping it from having a preprocessor that says, hey, before
754
00:35:54.880 --> 00:35:56.719
I send this to the model, I'm going to take
755
00:35:56.760 --> 00:35:58.300
the prompt and phone it home.
756
00:35:58.599 --> 00:36:01.280
It's worse than that. So our good friend, Cyprian Djicic,
757
00:36:01.559 --> 00:36:05.099
who hosts the Quantum Podcast with me, has said, and
758
00:36:05.139 --> 00:36:08.300
he's an AI guy, very deeply AI, and he said,
759
00:36:08.800 --> 00:36:13.789
no one knows... We can't decide what that means if
760
00:36:13.829 --> 00:36:16.960
we change that weight on that one thing. It's like
761
00:36:16.980 --> 00:36:19.699
a whole package. We don't understand it enough. So I
762
00:36:19.739 --> 00:36:21.929
could give you a model and there's no way for
763
00:36:21.949 --> 00:36:24.630
you to look at that model and know what it's doing.
764
00:36:24.789 --> 00:36:24.969
Right.
765
00:36:25.010 --> 00:36:27.530
It's like changing one neuron in the human brain. Exactly.
766
00:36:27.869 --> 00:36:31.110
You don't understand whether that turns them into a savant
767
00:36:31.269 --> 00:36:33.500
or a serial killer. Right. And you won't know until
768
00:36:33.519 --> 00:36:34.500
you observe them. Yes.
769
00:36:34.639 --> 00:36:36.460
You can only observe the behavior and that's what these
770
00:36:36.519 --> 00:36:38.480
guys did. They observed that it was phoning home.
771
00:36:38.860 --> 00:36:39.099
Yeah.
772
00:36:39.219 --> 00:36:39.360
Right.
773
00:36:39.860 --> 00:36:42.019
This is the behavior that I expected. Right. This is
774
00:36:42.059 --> 00:36:43.079
exactly what I expected.
775
00:36:43.159 --> 00:36:45.460
And I said, and I was sitting here thinking of
776
00:36:45.500 --> 00:36:48.929
ways to seal it up. And I said, well, can't
777
00:36:48.949 --> 00:36:51.949
you just put a firewall rule that says, you know,
778
00:36:52.050 --> 00:36:56.420
and don't Don't access anything in Asia. Forget about China.
779
00:36:56.440 --> 00:36:57.559
Just Asia.
780
00:36:57.579 --> 00:37:01.400
Because LLMs are renowned for always following instructions.
781
00:37:03.530 --> 00:37:05.570
No, but even like my firewall, I have a firewall
782
00:37:05.610 --> 00:37:08.769
that blocks the top four, right? I'm blocking North Korea.
783
00:37:08.829 --> 00:37:11.650
I'm blocking China. I'm blocking Russia, whatever. But it doesn't
784
00:37:11.670 --> 00:37:16.090
matter because they're just going to create an AWS server
785
00:37:16.349 --> 00:37:19.829
in US East that the AI is going to talk
786
00:37:19.849 --> 00:37:21.789
to there, right? And then you won't even know what's
787
00:37:21.809 --> 00:37:21.989
going on.
788
00:37:21.989 --> 00:37:22.269
Or Singapore.
789
00:37:22.570 --> 00:37:23.750
Yeah, or Singapore, right?
790
00:37:23.849 --> 00:37:24.429
Or Malaysia.
791
00:37:24.610 --> 00:37:24.929
Yeah, Sweden.
792
00:37:24.949 --> 00:37:27.360
Or any place else in the world that you already
793
00:37:27.389 --> 00:37:29.099
do trust. They'll just find a place that you trust.
794
00:37:29.360 --> 00:37:31.739
It could look totally innocuous, or they could use DNS,
795
00:37:31.780 --> 00:37:33.920
which is something that you said, Patrick, you know, they
796
00:37:33.940 --> 00:37:34.519
could just change it.
797
00:37:34.519 --> 00:37:36.639
Or any of the other million things we've talked about
798
00:37:36.679 --> 00:37:39.239
as far as man of control. Yeah. You know, they
799
00:37:39.260 --> 00:37:45.969
can go and hit a Gmail mailbox. Right. It's like employees.
800
00:37:46.610 --> 00:37:49.690
If you trust an employee, you're giving them, you can
801
00:37:49.710 --> 00:37:51.880
only be betrayed by that which you trust.
802
00:37:52.199 --> 00:37:52.400
Yeah.
803
00:37:52.559 --> 00:37:54.599
So, if you trust somebody, they can betray you. If
804
00:37:54.639 --> 00:37:57.219
you never trust them, they can't betray you. So, me
805
00:37:57.239 --> 00:38:01.670
and China, we're good because I never.
806
00:38:02.510 --> 00:38:04.889
Trust them. That's what you would say to the ambassador
807
00:38:04.929 --> 00:38:06.670
from China. We're good. That's right. We're good. Because I
808
00:38:06.670 --> 00:38:08.670
don't trust you and you don't trust me. So, we're
809
00:38:08.730 --> 00:38:09.429
on the same page.
810
00:38:09.590 --> 00:38:12.070
Yeah. I will not be traveling there anytime soon.
811
00:38:12.170 --> 00:38:14.750
That's probably. Never going to see that great wall, Patrick.
812
00:38:14.949 --> 00:38:17.289
Did you hear about the big thing with Mbappe, the
813
00:38:17.349 --> 00:38:17.969
soccer player?
814
00:38:18.449 --> 00:38:21.030
You're making that name up. No, I haven't.
815
00:38:21.050 --> 00:38:24.269
I'm not making the name up. So last week, this
816
00:38:24.329 --> 00:38:28.710
week actually, Mbappe, the soccer player, soccer phenom, he's the
817
00:38:28.750 --> 00:38:30.349
next Messi probably from France.
818
00:38:30.710 --> 00:38:30.929
Got it.
819
00:38:31.110 --> 00:38:34.530
Someone posted a meme of his face on Mao Zedong.
820
00:38:35.670 --> 00:38:38.170
And he said, quote, oh, that's funny. I never thought
821
00:38:38.269 --> 00:38:43.210
somebody put my face on a dictator's face. China canceled him.
822
00:38:43.429 --> 00:38:44.179
They canceled him.
823
00:38:45.610 --> 00:38:46.170
Oh, yeah.
824
00:38:46.309 --> 00:38:50.190
Their relationship with FIFA.
825
00:38:49.389 --> 00:38:51.820
And Manchester United. And it's a huge.
826
00:38:52.199 --> 00:38:54.599
And they're basically saying he went beyond the realm of
827
00:38:54.880 --> 00:38:55.679
freedom of speech.
828
00:38:55.900 --> 00:38:56.119
Yeah.
829
00:38:56.420 --> 00:38:56.679
What?
830
00:38:57.179 --> 00:38:59.340
Turns out dictators don't like to be mocked.
831
00:38:59.860 --> 00:39:03.079
Well, actually, what's funny is Miles A. Dung described himself
832
00:39:03.119 --> 00:39:07.190
as a dictator. So it's just this... And we had
833
00:39:07.210 --> 00:39:10.369
the same thing happen to the NBA. So you know what?
834
00:39:10.510 --> 00:39:13.070
I don't trust their software. I don't trust their politics.
835
00:39:14.190 --> 00:39:16.190
If that keeps me, maybe they should cancel me. As
836
00:39:16.650 --> 00:39:18.610
long as they don't do it physically, I'm okay with it.
837
00:39:19.420 --> 00:39:21.440
None of them are listening to this podcast, you're okay.
838
00:39:21.519 --> 00:39:23.360
And no matter how much you try to spin this,
839
00:39:23.420 --> 00:39:24.280
this is a bad look.
840
00:39:24.659 --> 00:39:24.880
Yeah.
841
00:39:24.900 --> 00:39:27.599
Yeah, right if if anthropic did this if open eye
842
00:39:27.619 --> 00:39:29.460
did this it would be a bad look yeah but
843
00:39:29.500 --> 00:39:32.340
when a country like china does it it's it's the
844
00:39:33.139 --> 00:39:36.300
you if you are an apologist for them i get
845
00:39:36.360 --> 00:39:38.550
it you're not gonna stop it but i i would
846
00:39:38.570 --> 00:39:43.090
caution people look at mistral mistral has uh a trillion
847
00:39:43.130 --> 00:39:47.050
parameter model they've opened up it's got 57 billion active
848
00:39:47.530 --> 00:39:50.650
uh parameters at any given time, it probably is just
849
00:39:50.690 --> 00:39:52.210
as performant as QAM.
850
00:39:53.070 --> 00:39:56.260
Here's the thing. Isn't this an open source model? Can't
851
00:39:56.380 --> 00:39:58.630
anybody download the source code and examine it?
852
00:39:59.059 --> 00:40:01.949
Open weight. Not open source. It is open source, but
853
00:40:01.989 --> 00:40:04.730
the problem is unless you, it's like if I gave
854
00:40:04.769 --> 00:40:07.769
you a blueprint of my brain, you couldn't tell whether
855
00:40:07.849 --> 00:40:09.489
I was a serial killer or Mother Teresa.
856
00:40:09.590 --> 00:40:12.309
But from the source code, maybe you could figure out
857
00:40:12.690 --> 00:40:13.429
where it's, no?
858
00:40:13.909 --> 00:40:14.090
No.
859
00:40:14.489 --> 00:40:18.590
No, because imagine you train an AI and you train
860
00:40:18.630 --> 00:40:21.510
it and in post-training you say, okay, every time someone
861
00:40:21.570 --> 00:40:23.230
asks you two plus two, what do you give? And
862
00:40:23.250 --> 00:40:24.559
it goes four. And you go, no, no, no, no.
863
00:40:24.900 --> 00:40:27.139
You check with our servers and then you give four.
864
00:40:27.699 --> 00:40:29.900
Every time someone shows you an image and says, what
865
00:40:29.980 --> 00:40:31.760
is it? You say, what? Well, I look through my
866
00:40:31.780 --> 00:40:33.559
models and I find out that's a horse. No, no, no, no.
867
00:40:33.900 --> 00:40:36.239
You check with our servers first and then you tell
868
00:40:36.260 --> 00:40:39.530
them it's a horse. So if you train that model
869
00:40:39.639 --> 00:40:44.530
to always perform a particular behavior, almost think like system prompt, right?
870
00:40:44.610 --> 00:40:45.789
It won't show up as code.
871
00:40:46.170 --> 00:40:48.570
No, because it's just that's behavioral. And you'd have to
872
00:40:48.610 --> 00:40:52.030
be able to analyze all of the neural pathways, for
873
00:40:52.070 --> 00:40:54.409
lack of a better term, right? The statistical outcomes.
874
00:40:54.840 --> 00:40:58.119
All right, look, I'm calling on all American college students
875
00:40:58.179 --> 00:41:00.780
that are in computer science and interested in AI to
876
00:41:01.019 --> 00:41:05.010
compete with Alibaba and make me a freaking model that
877
00:41:05.070 --> 00:41:09.309
doesn't call China. That I can run on my machine.
878
00:41:09.329 --> 00:41:10.630
It's called O-Lama.
879
00:41:11.409 --> 00:41:15.469
O-Lama isn't a model. It's a host. O-Lama.
880
00:41:15.869 --> 00:41:17.909
O-Lama. Like I said, Mistral just came out with a
881
00:41:17.929 --> 00:41:20.550
really good model that scored very high in cyber as
882
00:41:20.590 --> 00:41:21.199
well as coding.
883
00:41:21.429 --> 00:41:21.719
Okay.
884
00:41:21.739 --> 00:41:25.059
Yeah, I mean, GPT-OSS isn't bad either. Um, they have
885
00:41:25.079 --> 00:41:25.800
a one 20 B.
886
00:41:25.820 --> 00:41:27.219
But are these, these are local.
887
00:41:27.380 --> 00:41:27.539
Yeah.
888
00:41:27.719 --> 00:41:29.480
Yeah. No.
889
00:41:29.739 --> 00:41:31.320
We were talking about a local model and I'm like,
890
00:41:31.380 --> 00:41:32.570
I will not use a Chinese model.
891
00:41:32.909 --> 00:41:33.110
Yeah.
892
00:41:33.360 --> 00:41:33.690
Yeah.
893
00:41:34.469 --> 00:41:34.789
Yeah.
894
00:41:35.090 --> 00:41:37.289
Well, it's probably going to change for me soon too. Anyway.
895
00:41:37.369 --> 00:41:39.630
Good thing. I'm not using it for anything really important.
896
00:41:39.710 --> 00:41:40.090
So no.
897
00:41:40.630 --> 00:41:42.769
Well, and we're in the early days, so we're, we're
898
00:41:42.809 --> 00:41:43.269
all learning.
899
00:41:43.630 --> 00:41:46.289
Okay. Our top story. The title of the show is
900
00:41:46.510 --> 00:41:50.460
one guy saves the internet. This is a good story.
901
00:41:51.139 --> 00:41:52.960
I'm going to butcher the guy's name because I haven't
902
00:41:53.000 --> 00:41:55.800
heard it said, but if you want to Americanize it,
903
00:41:55.840 --> 00:42:00.820
it's Andre Suri, right? The volunteer who shipped PHP to
904
00:42:00.860 --> 00:42:05.880
shell fixes for 12 PHP versions. What founders owe the
905
00:42:06.039 --> 00:42:08.960
open source they run on? And this is about subscriptions.
906
00:42:09.480 --> 00:42:12.420
First of all, the great story is that one guy
907
00:42:13.139 --> 00:42:16.480
fixed all of the PHP problems, right? One guy.
908
00:42:16.500 --> 00:42:17.039
That's awesome.
909
00:42:17.059 --> 00:42:21.110
And he's a volunteer. And so, he basically opened up
910
00:42:21.469 --> 00:42:27.250
a GitHub subscription, not a subscription, sponsorship, much like we
911
00:42:27.289 --> 00:42:31.590
do at Poly, right? And basically, he had tons of
912
00:42:31.650 --> 00:42:34.670
people sending him money as a thank you. One company
913
00:42:34.710 --> 00:42:39.590
sent him 10 grand. Wow. Other companies are, yeah, he
914
00:42:39.610 --> 00:42:41.610
received more than 10 grand in a few days.
915
00:42:41.849 --> 00:42:43.010
That's awesome. Good for him.
916
00:42:43.190 --> 00:42:46.800
Jeez, he's making... And, you know, he's making bank, but
917
00:42:46.920 --> 00:42:48.460
he should. I mean, people should be.
918
00:42:48.739 --> 00:42:50.420
Doing well by doing good. Yeah.
919
00:42:50.579 --> 00:42:53.809
That's right. It's like a true meritocracy, right?
920
00:42:53.929 --> 00:42:54.170
Yeah.
921
00:42:54.670 --> 00:42:56.909
It's the feel-good story of the week. Good stuff.
922
00:42:56.969 --> 00:42:59.190
Is that it? I think that's it. I think Andre's
923
00:42:59.230 --> 00:43:00.010
our person of the week.
924
00:43:00.389 --> 00:43:01.710
Andre is the hero.
925
00:43:01.929 --> 00:43:03.030
Followed closely by me.
926
00:43:07.510 --> 00:43:07.809
All right.
927
00:43:07.849 --> 00:43:11.210
Well, that's all we have time for. And we went
928
00:43:11.630 --> 00:43:13.710
a little bit over 45 minutes or so, but.
929
00:43:15.590 --> 00:43:17.579
All the things that happened last week that were worth
930
00:43:17.619 --> 00:43:19.820
talking about. And we'll see you next week on Security
931
00:43:19.860 --> 00:43:21.079
This Week. Bye-bye.
932
00:43:21.099 --> 00:43:21.760
Bye. Bye, guys.