UTSKRIFT 🔗
1
00:00:00,000 --> 00:00:09,600
Welcome to the Azure Security Podcast, where we discuss topics relating to security, privacy,
2
00:00:09,600 --> 00:00:13,360
reliability and compliance on the Microsoft Cloud Platform.
3
00:00:13,360 --> 00:00:15,760
Hey, everybody.
4
00:00:15,760 --> 00:00:18,460
Welcome to episode 98.
5
00:00:18,460 --> 00:00:23,560
This week, it's myself, Michael, with Gladys, Sarah and Mark are both away today.
6
00:00:23,560 --> 00:00:29,560
And our guest this week is David Weston, who's here to talk to us about a whole slew of things,
7
00:00:29,560 --> 00:00:35,840
really from secure future initiative to software development to various other topics that kind
8
00:00:35,840 --> 00:00:37,320
of relate to that.
9
00:00:37,320 --> 00:00:40,880
But before we get to our guest, let's take a little lap around the news.
10
00:00:40,880 --> 00:00:42,840
Gladys, why don't you kick things off?
11
00:00:42,840 --> 00:00:44,440
Hi, everyone.
12
00:00:44,440 --> 00:00:51,400
I have been really busy with many other internal employees at Microsoft working on the secure
13
00:00:51,400 --> 00:00:54,800
future initiative or SFI.
14
00:00:54,800 --> 00:01:00,560
We will talk in a little bit about SFI, but first, let's talk about a public preview on
15
00:01:00,560 --> 00:01:07,000
Microsoft Defender for Cloud Integration with Microsoft Copilot for Security.
16
00:01:07,000 --> 00:01:13,800
This integration was first announced in MSBuild and now launched on June 10th.
17
00:01:13,800 --> 00:01:20,520
Defender for Cloud, as you may know, stands for the first cloud native application protection
18
00:01:20,520 --> 00:01:27,560
platform or CNAP, which is a solution that not only helps address posture related questions,
19
00:01:27,560 --> 00:01:35,120
but also assess security admin in understanding the environment, remediating issues and mitigating
20
00:01:35,120 --> 00:01:39,880
risk with AI generated actions.
21
00:01:39,880 --> 00:01:45,840
It includes a step by step guide with ready to execute scripts and pull requests for fixing
22
00:01:45,840 --> 00:01:48,080
vulnerabilities in code.
23
00:01:48,080 --> 00:01:56,120
There is also an assistant in identifying the appropriate resources, the owners and
24
00:01:56,120 --> 00:01:59,600
a developer for task delegation.
25
00:01:59,600 --> 00:02:07,920
This makes it easier for junior related team members to perform tasks like experience security
26
00:02:07,920 --> 00:02:08,920
administrators.
27
00:02:08,920 --> 00:02:12,600
It's definitely like a teaching tool.
28
00:02:12,600 --> 00:02:19,840
There's several free training videos, so please visit our site to see the links.
29
00:02:19,840 --> 00:02:27,320
Also I want to announce in public preview is a new SKU for Microsoft Azure Bastion Premium.
30
00:02:27,320 --> 00:02:34,440
This service is aimed for customers that handle high sensitivity virtual machines workloads.
31
00:02:34,440 --> 00:02:40,840
It allows customers to perform advanced recording, monitoring and auditing capabilities for sessions.
32
00:02:40,840 --> 00:02:44,240
This is something that many customers have been asking about.
33
00:02:44,240 --> 00:02:50,480
We finally have released this needed capability.
34
00:02:50,480 --> 00:02:51,920
I have one news item.
35
00:02:51,920 --> 00:02:57,440
It's an interesting one, not because of the product that it represents, but because of
36
00:02:57,440 --> 00:03:00,920
the overarching rationale.
37
00:03:00,920 --> 00:03:02,680
I'm going to read this verbatim.
38
00:03:02,680 --> 00:03:09,400
Log search alert rules using link storage will require using a managed identity starting
39
00:03:09,400 --> 00:03:14,880
July 2024, which is not far away.
40
00:03:14,880 --> 00:03:19,760
On the surface, this doesn't sound overly exciting, but in actual fact it's very important.
41
00:03:19,760 --> 00:03:27,480
We, Microsoft and certainly Azure specifically, are moving away from using credentials for
42
00:03:27,480 --> 00:03:29,240
basically anything.
43
00:03:29,240 --> 00:03:35,400
We know it's a long journey, but the point is we need to get credentials out of the environment.
44
00:03:35,400 --> 00:03:39,400
That includes not just us, but also customers of Azure.
45
00:03:39,400 --> 00:03:45,640
We need to move away from using usernames and passwords that get embedded somewhere,
46
00:03:45,640 --> 00:03:47,720
or SAS tokens, those kinds of things.
47
00:03:47,720 --> 00:03:49,000
We need to get away from those.
48
00:03:49,000 --> 00:03:52,440
The reason is because that's what the attackers are doing.
49
00:03:52,440 --> 00:03:56,920
They're getting hold of those credentials and that helps them move laterally in the
50
00:03:56,920 --> 00:03:57,920
environment.
51
00:03:57,920 --> 00:04:02,000
If the credentials are not there, then they can't get them.
52
00:04:02,000 --> 00:04:10,240
The way we solve that, certainly for client authentication in Azure, is to use enter ID.
53
00:04:10,240 --> 00:04:13,520
What that means for processes is managed identities.
54
00:04:13,520 --> 00:04:16,440
You really have two options, system managed identities and user managed identities.
55
00:04:16,440 --> 00:04:20,360
I'm going to provide a link in the notes.
56
00:04:20,360 --> 00:04:24,680
I really think everyone who's listening to this podcast really needs to dig deep and
57
00:04:24,680 --> 00:04:28,080
understand managed identities and why they're so important.
58
00:04:28,080 --> 00:04:32,120
Ultimately, there is no credential there and that's incredibly important from a client
59
00:04:32,120 --> 00:04:36,960
authentication perspective, but also from a, the attackers can't get the credential
60
00:04:36,960 --> 00:04:38,920
because it's not there.
61
00:04:38,920 --> 00:04:40,480
That's the most important part.
62
00:04:40,480 --> 00:04:49,200
Again, this story isn't that exciting, but the overarching rationale is critically important.
63
00:04:49,200 --> 00:04:50,760
That's the news
Load full transcript
out the way.
64
00:04:50,760 --> 00:04:52,840
Now let's turn our attention to our guest.
65
00:04:52,840 --> 00:04:57,240
As I mentioned at the top, our guest this week is David Weston, who's here to talk to
66
00:04:57,240 --> 00:05:01,200
us about a whole myriad of things, especially developer related stuff.
67
00:05:01,200 --> 00:05:04,120
David, thank you so much for joining us this week.
68
00:05:04,120 --> 00:05:07,080
We'd like to take a moment and introduce yourself to our listeners.
69
00:05:07,080 --> 00:05:08,080
Yeah, absolutely.
70
00:05:08,080 --> 00:05:13,640
I'm Dave Dwizzle Weston, as many of you who follow me on Twitter will know me as.
71
00:05:13,640 --> 00:05:17,840
I'm a vice president of operating system security here at Microsoft.
72
00:05:17,840 --> 00:05:22,680
I actually have two jobs, but I just get one paycheck, but my boss likes to point out it's
73
00:05:22,680 --> 00:05:25,160
a large paycheck, so I don't complain.
74
00:05:25,160 --> 00:05:28,880
I do building of security in the operating system.
75
00:05:28,880 --> 00:05:35,200
Everything from crypto to authentication, code signing, mitigations, you name it.
76
00:05:35,200 --> 00:05:38,880
I work on a team of builders and I lead them.
77
00:05:38,880 --> 00:05:43,720
One of the things that folks may not know is not only do we build many different flavors
78
00:05:43,720 --> 00:05:48,800
of windows from like Xbox to the Windows 11 you know and love to the Azure host, but we
79
00:05:48,800 --> 00:05:52,640
also build Linux distributions.
80
00:05:52,640 --> 00:05:58,040
A good example that is Azure sphere, which is our IOT project, which I hope lead has
81
00:05:58,040 --> 00:06:00,520
some flavor of a Linux distribution.
82
00:06:00,520 --> 00:06:06,720
We also use Linux now directly on Azure host as part of Azure boost, and we actually have
83
00:06:06,720 --> 00:06:12,960
a Linux distribution, RPM based that anyone can download called Azure Linux, and we actually
84
00:06:12,960 --> 00:06:15,560
do a fair bit of work in upstreaming on that.
85
00:06:15,560 --> 00:06:19,920
Besides building the operating systems, I also lead a penetration testing team that
86
00:06:19,920 --> 00:06:26,880
does things like code reviews, implementing the secure future initiative or SFI, SDL,
87
00:06:26,880 --> 00:06:28,440
red teaming, you name it.
88
00:06:28,440 --> 00:06:32,920
So we kind of have this really cool culture of people building security products, but
89
00:06:32,920 --> 00:06:37,400
also a lot of offensive expertise, and I think that really helps to make a better product.
90
00:06:37,400 --> 00:06:43,400
All right, so let's start off with something you mentioned at the top, which is Dwizzle
91
00:06:43,400 --> 00:06:45,160
with three Z's.
92
00:06:45,160 --> 00:06:47,320
So where on earth does Dwizzle come from?
93
00:06:47,320 --> 00:06:52,760
Yeah, you know, when you red team and you find good bugs, you often entertain yourself
94
00:06:52,760 --> 00:06:53,760
with those bugs.
95
00:06:53,760 --> 00:06:56,800
So this is years and years back.
96
00:06:56,800 --> 00:07:00,500
As Michael will know, many of you will know, you know, there's a way to sort of present
97
00:07:00,500 --> 00:07:06,120
a headshot or profile picture in Outlook in the exchange client that actually used to
98
00:07:06,120 --> 00:07:09,240
be held on a SharePoint server way back when.
99
00:07:09,240 --> 00:07:13,560
And we actually found an arbitrary file upload vulnerability years back that would allow
100
00:07:13,560 --> 00:07:19,400
you to if you knew the picture name, you could overwrite anyone's photo.
101
00:07:19,400 --> 00:07:24,800
And so of course, I immediately rather than reporting that bug, used it to draw eye patches
102
00:07:24,800 --> 00:07:28,080
on my red team colleagues back in those days.
103
00:07:28,080 --> 00:07:30,320
And so of course, we got it fixed.
104
00:07:30,320 --> 00:07:34,820
But those eye patches stayed in the Outlook client for a while.
105
00:07:34,820 --> 00:07:41,760
And eventually, some folks found a way to change the presentation name in exchange and
106
00:07:41,760 --> 00:07:45,160
they paid me back by changing my name to Dwizzle.
107
00:07:45,160 --> 00:07:46,880
And I actually found it was kind of cool.
108
00:07:46,880 --> 00:07:49,120
Pretty soon, execs were addressing me by it.
109
00:07:49,120 --> 00:07:51,200
And then I was branded and couldn't get rid of it.
110
00:07:51,200 --> 00:07:54,320
So that's where it stands.
111
00:07:54,320 --> 00:07:56,240
And you know, along the way, there's been some other hijinks.
112
00:07:56,240 --> 00:08:02,160
Like I've done some reports in some pretty prestigious journals where on a dare or a
113
00:08:02,160 --> 00:08:07,000
bet loss, you know, I was attributed to Dwizzle of the M dollar, things like that.
114
00:08:07,000 --> 00:08:09,960
So yeah, the history goes back a long ways.
115
00:08:09,960 --> 00:08:13,880
And most of the red team folks here at Microsoft have pretty similar stories.
116
00:08:13,880 --> 00:08:19,320
M dollar, man, I haven't seen that as an abbreviation for Microsoft for a long time.
117
00:08:19,320 --> 00:08:20,320
That's funny.
118
00:08:20,320 --> 00:08:21,320
That's funny.
119
00:08:21,320 --> 00:08:22,320
The M dollar.
120
00:08:22,320 --> 00:08:23,320
That's right.
121
00:08:23,320 --> 00:08:26,240
I'm dating myself for sure.
122
00:08:26,240 --> 00:08:32,120
Microsoft has been talking about the Secure Future Initiative or SFI.
123
00:08:32,120 --> 00:08:37,920
And many of us have been working intensively in it, as I mentioned earlier.
124
00:08:37,920 --> 00:08:42,320
Can you talk a little bit about it and your role in it?
125
00:08:42,320 --> 00:08:47,480
Yeah, the Secure Future Initiative, the way I generally look at it is, you know, a cultural
126
00:08:47,480 --> 00:08:50,400
change and a prioritization across Microsoft.
127
00:08:50,400 --> 00:08:51,640
Security becomes job one.
128
00:08:51,640 -->
00:08:56,000
You know, I definitely think initiatives like Trustworthy Computing and SDL have certainly
129
00:08:56,000 --> 00:08:57,480
done that in the past.
130
00:08:57,480 --> 00:08:59,880
I look at SFI as even more all encompassing.
131
00:08:59,880 --> 00:09:05,800
You know, SDL and things like that talk specifically about how we're going to be secure by construction.
132
00:09:05,800 --> 00:09:12,240
But SFI includes things like operational security, mindset, training, and I would say a secure
133
00:09:12,240 --> 00:09:14,840
by default ethos.
134
00:09:14,840 --> 00:09:17,800
And so I think that I play sort of two major roles in that.
135
00:09:17,800 --> 00:09:22,960
The first is as someone building the operating system, I think SFI really helps and supports
136
00:09:22,960 --> 00:09:28,200
what I think the security team's objective always is, which is to prioritize security.
137
00:09:28,200 --> 00:09:31,080
You know, there's a lot of talk of that online, especially in the security community, which
138
00:09:31,080 --> 00:09:35,440
is sometimes a little cavalier about how difficult that is.
139
00:09:35,440 --> 00:09:38,360
At the end of the day, when you're building an operating system product, you know, people
140
00:09:38,360 --> 00:09:42,720
are not going to buy that if their software can't run or if it's too slow.
141
00:09:42,720 --> 00:09:47,300
So a lot of what we do is balancing secure by default, but also figuring out from an
142
00:09:47,300 --> 00:09:51,680
engineering perspective, how do we also balance compatibility, which is sometimes, you know,
143
00:09:51,680 --> 00:09:55,260
the enemy of security and also maintaining performance.
144
00:09:55,260 --> 00:10:00,200
And so one of the big steps my team has done is taken the step of turning on things like
145
00:10:00,200 --> 00:10:05,520
virtualization based security by default, BitLocker by default, code signing and code integrity
146
00:10:05,520 --> 00:10:08,520
by default with features like smart app control.
147
00:10:08,520 --> 00:10:09,920
And I think, Michael, you'll like this.
148
00:10:09,920 --> 00:10:10,920
We're really close.
149
00:10:10,920 --> 00:10:14,680
Hopefully in the near future, you'll hear more from me on this, but we're really close
150
00:10:14,680 --> 00:10:19,480
to removing admin by default and having just in time admin become the default in Windows.
151
00:10:19,480 --> 00:10:24,360
So I think that's a good, you know, example of where SFI is being put into action.
152
00:10:24,360 --> 00:10:28,440
We're empowered to do things that maybe previously we would have optimized for performance or
153
00:10:28,440 --> 00:10:29,440
compatibility.
154
00:10:29,440 --> 00:10:35,100
But outside of that, my teams do a lot to secure operating system related services as
155
00:10:35,100 --> 00:10:37,000
well as security features.
156
00:10:37,000 --> 00:10:42,520
So of course, implementing the SDL from the SDL you know and love, which is threat modeling,
157
00:10:42,520 --> 00:10:48,920
static analysis, fuzzing, security bug bars and response, but also taking that up a notch,
158
00:10:48,920 --> 00:10:53,200
you know, redesigning, I would say features to take advantage of kind of modern security
159
00:10:53,200 --> 00:10:54,200
boundaries.
160
00:10:54,200 --> 00:10:58,480
So in many cases, we're getting very involved at the design level to use things like the
161
00:10:58,480 --> 00:11:01,760
virtualization based security boundary or hardware.
162
00:11:01,760 --> 00:11:05,480
We also get involved pretty deeply in the operational security.
163
00:11:05,480 --> 00:11:10,320
So things like testing app specific defenses and making sure we're building bespoke detections
164
00:11:10,320 --> 00:11:11,780
through red teaming.
165
00:11:11,780 --> 00:11:17,360
So across the board from design to implementation, you know, my team gets pretty involved there,
166
00:11:17,360 --> 00:11:21,080
which means I spend a lot of time sort of helping guide them towards a strategy that
167
00:11:21,080 --> 00:11:23,760
can execute well on that.
168
00:11:23,760 --> 00:11:27,760
So talking about compatibility, I noticed just recently there was an announcement made
169
00:11:27,760 --> 00:11:33,800
by the Windows Server 2025 team that NTLM was finally going bye bye.
170
00:11:33,800 --> 00:11:37,120
But also mail slots.
171
00:11:37,120 --> 00:11:43,200
I think the last, if I remember correctly, didn't mail slots like first start in landman
172
00:11:43,200 --> 00:11:45,480
in the land manager on OS2?
173
00:11:45,480 --> 00:11:47,280
I think it's all the way back to OS2.
174
00:11:47,280 --> 00:11:48,280
Yes, OS2.
175
00:11:48,280 --> 00:11:49,280
Yeah.
176
00:11:49,280 --> 00:11:51,000
So there's a whole bunch of other stuff that's being deprecated.
177
00:11:51,000 --> 00:11:55,720
I mean, you know, I imagine the code behind that is probably old and crusty and you know,
178
00:11:55,720 --> 00:11:57,840
it's just just time to deprecate it.
179
00:11:57,840 --> 00:12:02,040
Probably very old C, barely C++ code.
180
00:12:02,040 --> 00:12:06,760
And I'm sure it's been reviewed and fuzz tested over the years, but by the same token, it's
181
00:12:06,760 --> 00:12:08,520
probably just just old and crusty, right?
182
00:12:08,520 --> 00:12:14,720
So it's time to sort of, I don't even know, does any feature actually even use mail slots?
183
00:12:14,720 --> 00:12:19,200
I don't know if anything inboxes us, but certainly, you know, you wouldn't be surprised to hear
184
00:12:19,200 --> 00:12:23,200
this, but there's many, many crusty old applications that people rely on Windows.
185
00:12:23,200 --> 00:12:27,720
And so there are certainly third party applications that continue to do that.
186
00:12:27,720 --> 00:12:32,680
You know, but I think attack surface removal is a big sort of implied part of secure by
187
00:12:32,680 --> 00:12:33,680
design.
188
00:12:33,680 --> 00:12:37,880
For example, you know,
things like NTLM, we know at the end of the day, there are limitations
189
00:12:37,880 --> 00:12:42,040
to the NTLM as a protocol and as the standard in terms of security.
190
00:12:42,040 --> 00:12:45,600
And so when you talk about security by default in that particular area, I think the best
191
00:12:45,600 --> 00:12:48,200
thing we can do is, you know, disable it.
192
00:12:48,200 --> 00:12:50,160
I think that goes for a lot of attack surface.
193
00:12:50,160 --> 00:12:54,600
There are many, many, I would say perennial features where we've had security challenges
194
00:12:54,600 --> 00:12:56,920
around them that we're really looking at redesigning.
195
00:12:56,920 --> 00:13:04,400
You know, I'd throw Win32k out there, NTLM, and there are many other surfaces that, you
196
00:13:04,400 --> 00:13:10,920
know, we continue to look at to try to either remove or fundamentally redesign.
197
00:13:10,920 --> 00:13:15,400
And that's actually the exciting part of Windows right now, I think, is there's a ton of energy
198
00:13:15,400 --> 00:13:19,620
and a ton of support at the leadership level for really revolutionizing the guarantees
199
00:13:19,620 --> 00:13:20,620
that we can make.
200
00:13:20,620 --> 00:13:22,680
So that's always fun, right?
201
00:13:22,680 --> 00:13:27,160
When you think you kind of know what the limit is, I think we're in the midst of redefining
202
00:13:27,160 --> 00:13:30,920
it, which is an exciting time to work on an operating system at Microsoft.
203
00:13:30,920 --> 00:13:36,500
So why you mentioned redefining, perhaps even redesigning and even mentioned revolutionary
204
00:13:36,500 --> 00:13:37,500
in there.
205
00:13:37,500 --> 00:13:38,500
Rust.
206
00:13:38,500 --> 00:13:42,640
Tell me about, I mean, I'm a huge Rust fan.
207
00:13:42,640 --> 00:13:48,180
And that's coming from someone who's done C and C++ code for a long, long, long time,
208
00:13:48,180 --> 00:13:49,720
but I enjoy doing Rust.
209
00:13:49,720 --> 00:13:52,680
So tell me what's going on at Microsoft around Rust.
210
00:13:52,680 --> 00:13:53,680
Sure.
211
00:13:53,680 --> 00:13:57,280
So the first thing I'll say is I think Microsoft actually has a long history with, I'll call
212
00:13:57,280 --> 00:13:59,040
it type and memory safety.
213
00:13:59,040 --> 00:14:03,360
Michael, you'll remember this, but big parts of Vista were actually originally designed
214
00:14:03,360 --> 00:14:06,240
in C sharp and dot net.
215
00:14:06,240 --> 00:14:10,520
And one of the reasons was to achieve, you know, type and memory safety.
216
00:14:10,520 --> 00:14:16,000
We actually had a pretty long running experimental operating system called Midori that used a
217
00:14:16,000 --> 00:14:22,460
version of dot net that we called system C sharp that had pretty strong security guarantees.
218
00:14:22,460 --> 00:14:30,340
And even in products like Azure sphere, we had a subset of C called safer C that required
219
00:14:30,340 --> 00:14:32,140
some things around
spatial safety, et cetera.
220
00:14:32,140 --> 00:14:34,960
So that's been a long running goal.
221
00:14:34,960 --> 00:14:39,380
Traditionally the problem is, well, you know, languages that offer things like temporal
222
00:14:39,380 --> 00:14:44,200
safety and spatial safety and things like garbage collection that come along with that
223
00:14:44,200 --> 00:14:49,600
just usually incongruent with the performance and efficiency needs of a low level operating
224
00:14:49,600 --> 00:14:50,600
system.
225
00:14:50,600 --> 00:14:53,840
You know, having a garbage collector, for example, in your bootloader, things like that,
226
00:14:53,840 --> 00:14:55,080
generally incongruent.
227
00:14:55,080 --> 00:15:01,640
So I think really what Rust did and to an extent Golang and other languages like Swift
228
00:15:01,640 --> 00:15:05,900
have really brought about, I think, a revolution in systems level languages.
229
00:15:05,900 --> 00:15:09,980
So languages that are appropriate from an efficiency and performance standpoint to use
230
00:15:09,980 --> 00:15:14,040
an operating system construction while also having memory safety.
231
00:15:14,040 --> 00:15:18,960
And so I think that people understand the elegance and the approach of that and the
232
00:15:18,960 --> 00:15:24,120
benefits it gives you when this sort of secure by construction, especially deterministically.
233
00:15:24,120 --> 00:15:26,280
And so I think it's caught the eye of folks.
234
00:15:26,280 --> 00:15:30,640
You know, Mozilla originally designed this language as a way to secure their browser.
235
00:15:30,640 --> 00:15:32,240
That makes a lot of sense in that.
236
00:15:32,240 --> 00:15:36,700
But I think the community as a whole has been waiting for something like this.
237
00:15:36,700 --> 00:15:40,520
So we've seen a lot of folks gravitating towards it, including at Microsoft.
238
00:15:40,520 --> 00:15:44,880
And really just we haven't had this opportunity for a very long time to get the performance
239
00:15:44,880 --> 00:15:47,200
we need out of our memory safe language.
240
00:15:47,200 --> 00:15:52,480
The thing I'll say about Rust in particular, why it's moving so fast is while the security
241
00:15:52,480 --> 00:15:57,840
folks you can call us wonks maybe have wanted this for a long time, what we're actually
242
00:15:57,840 --> 00:16:04,760
seeing what I'm observing is kind of a groundswell of support from your average individual contributor
243
00:16:04,760 --> 00:16:05,760
developer.
244
00:16:05,760 --> 00:16:06,820
They love Rust.
245
00:16:06,820 --> 00:16:09,440
They want to write code that's secure by construction.
246
00:16:09,440 --> 00:16:11,220
They're very interested in the language.
247
00:16:11,220 --> 00:16:15,920
So a lot of the changes are less Microsoft strategy, although I'm happy to talk about
248
00:16:15,920 --> 00:16:21,000
why it's on Microsoft strategy and more of the passionate, you know, boots on the ground
249
00:16:21,000 --> 00:16:25,240
devs are just saying, I want to do this and going and building really cool stuff with
250
00:16:25,240 --> 00:16:26,240
it.
251
00:16:26,240 --> 00:16:27,240
That's how we got this far.
252
00:16:27,240 --> 00:16:30,680
And so I want to make sure that those intrepid folks are really the ones that get the credit
253
00:16:30,680 --> 00:16:33,580
because I think that is the truth.
254
00:16:33,580 --> 00:16:39,280
So while while you're on a roll, I mean, why don't you sort of itemize a couple of things
255
00:16:39,280 --> 00:16:43,160
that are currently using rust and talk about some of the strategy.
256
00:16:43,160 --> 00:16:44,160
I think it's important.
257
00:16:44,160 --> 00:16:47,160
I mean, I think rust is an important language.
258
00:16:47,160 --> 00:16:49,320
It's definitely not going away.
259
00:16:49,320 --> 00:16:51,360
It's evolving rapidly.
260
00:16:51,360 --> 00:16:54,160
You could say that's good and bad.
261
00:16:54,160 --> 00:16:58,200
And then we can talk and then let's just talk about, you know, what's the role of C++ here?
262
00:16:58,200 --> 00:16:59,720
In fact, even C sharp as well, right?
263
00:16:59,720 --> 00:17:02,640
Because they're not languages that are going to go away.
264
00:17:02,640 --> 00:17:04,840
But I'm will just leave the button to laser.
265
00:17:04,840 --> 00:17:11,040
Yeah, so I think we got kind of our feet wet with some really important projects.
266
00:17:11,040 --> 00:17:15,960
So the first is I would say the two things that have maybe been the most attacked from
267
00:17:15,960 --> 00:17:19,960
a pure operating system component standpoint in terms of memory safety.
268
00:17:19,960 --> 00:17:25,440
And for those of you don't know, memory safety issues are about 70 to 75 percent of what
269
00:17:25,440 --> 00:17:30,240
we patch via MSRC issues and, you know, the updates you get through Windows updates.
270
00:17:30,240 --> 00:17:34,440
So just to set set expectations, about three quarters of the bugs Microsoft fix are in
271
00:17:34,440 --> 00:17:36,680
that memory safety category.
272
00:17:36,680 --> 00:17:40,740
And maybe the two most attacked components have been our font parsing.
273
00:17:40,740 --> 00:17:44,800
So the things that would parse, you know, TTF or OTF, especially on remote attack services
274
00:17:44,800 --> 00:17:50,720
like the browser or office clients, as well as Win32K.
275
00:17:50,720 --> 00:17:54,240
Win32K is our internal graphics component that has a long, long legacy.
276
00:17:54,240 --> 00:17:57,320
I think it was originally designed in the late 80s.
277
00:17:57,320 --> 00:18:02,440
And it's become frequently attacked because it offers a convenient escalation of privilege
278
00:18:02,440 --> 00:18:03,440
attack.
279
00:18:03,440 --> 00:18:08,200
So if you want to get out of a browser sandbox and Windows, you know, Win32K is often a common
280
00:18:08,200 --> 00:18:09,200
choice.
281
00:18:09,200 --> 00:18:13,200
And so we started Rust right in the place
where it would offer the most security value,
282
00:18:13,200 --> 00:18:21,640
which is taking the sort of modern WinApp SDK font parser that we call DirectWrite and
283
00:18:21,640 --> 00:18:23,960
converting that into Rust.
284
00:18:23,960 --> 00:18:25,840
And that gave us an idea.
285
00:18:25,840 --> 00:18:29,880
And so I did a presentation on this about a year ago, but it took roughly two to three
286
00:18:29,880 --> 00:18:32,380
months of a couple of developers time.
287
00:18:32,380 --> 00:18:35,520
And the interesting things that came out of it is that performance actually got better.
288
00:18:35,520 --> 00:18:36,520
Right.
289
00:18:36,520 --> 00:18:40,240
So I mentioned sort of at the top of my spiel is that traditional problem with memory safety
290
00:18:40,240 --> 00:18:46,520
is it was thought of as inferior in terms of efficiency and performance to C and C++.
291
00:18:46,520 --> 00:18:52,000
Well, our first experiment, although it's a small number, you know, small value, ultimately
292
00:18:52,000 --> 00:18:55,480
proved that you can write high performance Rust code.
293
00:18:55,480 --> 00:19:00,400
And so our second experiment, which was more learning about interop and less about replacing
294
00:19:00,400 --> 00:19:06,160
the whole component during performance was some GDI surfaces in Win32K.
295
00:19:06,160 --> 00:19:09,140
And there we were trying to learn about what is the extent of interop?
296
00:19:09,140 --> 00:19:14,880
Can we sort of slice out individual components of Win32K and make sure that those can work
297
00:19:14,880 --> 00:19:15,880
well?
298
00:19:15,880 --> 00:19:19,160
This is especially important because, you know, Microsoft Windows is compiled with the
299
00:19:19,160 --> 00:19:26,840
Visual C++ compiler or C compiler, while Rust's backend is actually LLVM and uses that backend.
300
00:19:26,840 --> 00:19:30,680
And so just making sure that we have, you know, ABI's that are going to work together,
301
00:19:30,680 --> 00:19:32,480
et cetera, is really important.
302
00:19:32,480 --> 00:19:35,360
And again, that was successful and is now shipped in Windows.
303
00:19:35,360 --> 00:19:39,000
So those were really the two projects that I don't know that they've gotten us to the
304
00:19:39,000 --> 00:19:42,560
north star in terms of memory safety, but got us to learn a lot.
305
00:19:42,560 --> 00:19:48,120
And so based on that, as part of the SF5, we have made a commitment moving forward to
306
00:19:48,120 --> 00:19:53,720
start to do parts of our trusted computing base or the most security critical new code
307
00:19:53,720 --> 00:19:55,340
will be written in Rust.
308
00:19:55,340 --> 00:19:57,860
And to achieve that goal, we have to do a few things.
309
00:19:57,860 --> 00:20:02,960
The first is we need a stable tool chain with compilers and other things that can kind of
310
00:20:02,960 --> 00:20:07,840
work in that large scale, incredible scale Windows environment.
311
00:20:07,840 --> 00:20:09,640
So you got to be able to debug with WinDebug.
312
00:20:09,640 --> 00:20:11,300
You have to have PDB files.
313
00:20:11,300 --> 00:20:14,540
You have to make sure code generation is stable.
314
00:20:14,540 --> 00:20:20,600
And so we announced that we were spending as a company around $10 million on Rust-based
315
00:20:20,600 --> 00:20:23,440
tools with Azure being the first target.
316
00:20:23,440 --> 00:20:26,920
And I'll talk a little bit about where we're using Rust in Azure to start.
317
00:20:26,920 --> 00:20:32,160
The second thing we did is we have donated a million dollars or in the process of donating
318
00:20:32,160 --> 00:20:36,480
to the Rust Foundation with really a goal of stabilizing those tools.
319
00:20:36,480 --> 00:20:41,440
Ideally, we'd like something equivalent to the Windows or sorry, the Linux long-term
320
00:20:41,440 --> 00:20:42,800
servicing branch.
321
00:20:42,800 --> 00:20:45,440
We want like sort of an LTS version of Rust.
322
00:20:45,440 --> 00:20:50,040
You can imagine calling Rust up in the middle of the Windows build is probably not ideal.
323
00:20:50,040 --> 00:20:56,840
So we want something that has a lifetime support of several years and we can base sort of future
324
00:20:56,840 --> 00:20:57,840
Windows off of that.
325
00:20:57,840 --> 00:21:02,600
So really the major thing that we've got to take care of right now is which is sort of
326
00:21:02,600 --> 00:21:07,680
in front of us writing large scale Rust in both Azure and Windows is getting those tools
327
00:21:07,680 --> 00:21:10,320
in that long-term servicing branch up.
328
00:21:10,320 --> 00:21:11,940
But we're hard at work at that.
329
00:21:11,940 --> 00:21:14,080
And so I expect a lot more announcements on that.
330
00:21:14,080 --> 00:21:18,660
And the other thing I wanted to touch on is Mark Resendovich, our CTO in Azure, he announced
331
00:21:18,660 --> 00:21:25,200
a commitment to moving towards Rust and again, a contingent on this tool chain.
332
00:21:25,200 --> 00:21:28,840
But one of the areas that we're spending a lot of time writing Rust is in a product called
333
00:21:28,840 --> 00:21:30,160
Azure Boost.
334
00:21:30,160 --> 00:21:34,840
Azure Boost is available now, but it's also sort of our future architecture for Azure.
335
00:21:34,840 --> 00:21:41,960
And that's where we are offloading more of the performant aspects of Azure Hosts onto
336
00:21:41,960 --> 00:21:46,880
specialized cards like SmartNICs and our FPGAs for storage.
337
00:21:46,880 --> 00:21:52,640
The goal there is twofold really to reduce jitter in a multi-tenant shared environment
338
00:21:52,640 --> 00:21:54,080
like an Azure Host.
339
00:21:54,080 --> 00:21:58,680
If your neighbor in another VM is taking a lot of IOPS, potentially that can impact your
340
00:21:58,680 --> 00:21:59,680
workload.
341
00:21:59,680 --> 00:22:04,240
So Boost eliminates that by doing more hardware offload, allowing sort of deterministic performance
342
00:22:04,240 --> 00:22:06,960
closer to what people would have on bare metal.
343
00:22:06,960 --> 00:22:09,120
But it also eliminates attack surface.
344
00:22:09,120 --> 00:22:13,520
So a lot of the network surface, a lot of the storage surface moves away from being
345
00:22:13,520 --> 00:22:18,080
software implemented to hardware emulated where you can use more of the hardware verification
346
00:22:18,080 --> 00:22:22,760
technologies, more deterministic secure by construction approaches that you would get
347
00:22:22,760 --> 00:22:26,120
in integrated circuit designs, et cetera.
348
00:22:26,120 --> 00:22:28,080
And so there's a huge benefit to that.
349
00:22:28,080 --> 00:22:33,560
Part of reducing jitter has also been moving more things, more agents that do things like
350
00:22:33,560 --> 00:22:37,880
network routing, et cetera, which are ultimately hypervisor attack surface or cloud critical
351
00:22:37,880 --> 00:22:39,420
attack surface.
352
00:22:39,420 --> 00:22:44,600
Off of sort of DOM 0 and Hyper-V, the traditional host attack surface, and onto what we call
353
00:22:44,600 --> 00:22:51,200
the Azure Boost control plane, which is a small embedded operating system running Linux.
354
00:22:51,200 --> 00:22:55,240
And more of that agentry is being re-implemented in Rust as a standard.
355
00:22:55,240 --> 00:23:01,240
So we are really betting on Rust as our future implementation path.
356
00:23:01,240 --> 00:23:04,660
In fact, if this blog is on reality, I can break some news here.
357
00:23:04,660 --> 00:23:09,840
We have a future VMM that we're using for Azure that will manage sort of Hyper-V, and
358
00:23:09,840 --> 00:23:12,640
that is being re-implemented in Rust and called Underhill.
359
00:23:12,640 --> 00:23:15,860
So if that blog isn't out yet, you'll see it very soon.
360
00:23:15,860 --> 00:23:20,800
So I can say without a doubt that we're betting heavily on Rust in Azure.
361
00:23:20,800 --> 00:23:25,040
It's a big part of our secure future initiative strategy.
362
00:23:25,040 --> 00:23:29,640
And the last thing I'll say is, you know, I've done, and my side sort of as a side project,
363
00:23:29,640 --> 00:23:34,120
I work with CISA, which is part of the DHS that looks after cybersecurity here in the
364
00:23:34,120 --> 00:23:35,360
United States.
365
00:23:35,360 --> 00:23:40,300
And I work on the technical advisory committee led by Jeff Moss, who founded DEF CON and
366
00:23:40,300 --> 00:23:41,560
Black Hat.
367
00:23:41,560 --> 00:23:48,200
And we wrote a paper recently on memory safety that's helped CISA kind of codify their secure
368
00:23:48,200 --> 00:23:50,280
by design portions.
369
00:23:50,280 --> 00:23:57,240
And so we're seeing a lot of momentum both from them as well as the White House's cybersecurity
370
00:23:57,240 --> 00:23:59,400
strategy department around memory safety.
371
00:23:59,400 --> 00:24:05,120
So I think Microsoft is really in line with kind of where the industry momentum is going.
372
00:24:05,120 --> 00:24:09,000
And it's been really exciting to see how fast we're moving with Rust and other memory safety
373
00:24:09,000 --> 00:24:30,040
technologies.
374
00:24:39,000 --> 00:24:49,440
Yeah, actually, Google has an amazing free public training on Rust.
375
00:24:49,440 --> 00:24:52,600
I don't have the link, but I'm sure we could find that to get you in the show notes.
376
00:24:52,600 --> 00:24:54,440
I've run through it and it's pretty awesome.
377
00:24:54,440 --> 00:24:57,640
There are also a number of free books on Rust.
378
00:24:57,640 --> 00:25:03,360
I think Rust documentation right off the rustlang.org is actually sort of a free ebook that you
379
00:25:03,360 --> 00:25:05,000
can print PDF.
380
00:25:05,000 --> 00:25:10,600
And between the course from Google and that book, I think you can get a very long way
381
00:25:10,600 --> 00:25:15,440
to implementing your first Rust programs at high quality.
382
00:25:15,440 --> 00:25:17,120
And both of those are completely free.
383
00:25:17,120 --> 00:25:19,760
Yeah, I want to back that up.
384
00:25:19,760 --> 00:25:24,720
I think the secret, whenever I'm learning a new programming language, one of the first
385
00:25:24,720 --> 00:25:27,840
things I do is set myself a project to build.
386
00:25:27,840 --> 00:25:28,840
Absolutely.
387
00:25:28,840 --> 00:25:31,200
99 times over 100, that's a web server.
388
00:25:31,200 --> 00:25:32,560
So I build a web server.
389
00:25:32,560 --> 00:25:33,560
That's a good idea.
390
00:25:33,560 --> 00:25:35,720
So you're doing network IO, right?
391
00:25:35,720 --> 00:25:43,160
And then you're fetching files and then you're caching and then you're doing multi-threaded
392
00:25:43,160 --> 00:25:47,520
and then for grins and giggles, you add TLS support on top of it.
393
00:25:47,520 --> 00:25:51,600
And don't get me wrong, I just do some of the simple verbs, get, put, post, and that's
394
00:25:51,600 --> 00:25:53,800
about it.
395
00:25:53,800 --> 00:25:59,840
But yeah, Rust is not a difficult language to learn.
396
00:25:59,840 --> 00:26:06,280
The only problem, I hate to say problem, but the only thing that you will, anyone who's
397
00:26:06,280 --> 00:26:10,540
learning Rust will end up pulling their hair out is the borrow checker.
398
00:26:10,540 --> 00:26:15,320
Until you get used to the borrow checker, you're going to be absolutely just punching
399
00:26:15,320 --> 00:26:18,040
your screen with your laptop.
400
00:26:18,040 --> 00:26:21,680
I found that, but once I understood the borrow checker and how it worked and why it was so
401
00:26:21,680 --> 00:26:27,560
useful and frankly unique to Rust, as far as I know, it's actually very-
402
00:26:27,560 --> 00:26:31,840
There's actually a borrow checker implemented in C++ now that's pretty interesting in the
403
00:26:31,840 --> 00:26:34,520
C++ compiler called Circle.
404
00:26:34,520 --> 00:26:35,520
Is that right?
405
00:26:35,520 --> 00:26:38,000
Is it part of a
standard or is it just some experimental?
406
00:26:38,000 --> 00:26:40,160
I believe it's not a standard.
407
00:26:40,160 --> 00:26:43,600
There's a fella called Sean Baxter who's got a really interesting Twitter.
408
00:26:43,600 --> 00:26:46,480
If you're interested in memory safety, I recommend following him.
409
00:26:46,480 --> 00:26:52,720
And he's got a very full featured compiler for C++ called Circle that has, I don't want
410
00:26:52,720 --> 00:26:58,840
to take words out of the mouth, but most of the memory safety features in Rust are actually
411
00:26:58,840 --> 00:27:02,520
supported with C++ semantics or the full language.
412
00:27:02,520 --> 00:27:04,840
That's been a pretty interesting thing to watch come together.
413
00:27:04,840 --> 00:27:05,840
Interesting.
414
00:27:05,840 --> 00:27:07,840
While we're on that topic then, so C++.
415
00:27:07,840 --> 00:27:13,400
You and I spoke about this some time ago and I was trying to say, hey, but if you're doing
416
00:27:13,400 --> 00:27:17,400
C++, what about modern C++, blah, blah, blah, blah, blah.
417
00:27:17,400 --> 00:27:21,240
You had a strong objection to that argument.
418
00:27:21,240 --> 00:27:23,760
Which is fine, which is completely fine.
419
00:27:23,760 --> 00:27:27,760
So let's go through that.
420
00:27:27,760 --> 00:27:31,200
Why shouldn't people seriously consider modern C++?
421
00:27:31,200 --> 00:27:37,960
The whole goal of modern C++ is that things like pointers and array offsets and so on
422
00:27:37,960 --> 00:27:41,720
are handled for you as opposed to you doing it by yourself.
423
00:27:41,720 --> 00:27:44,120
What's your opinion there?
424
00:27:44,120 --> 00:27:47,760
I think when we say memory safety, we're talking about completeness.
425
00:27:47,760 --> 00:27:53,400
So I don't have any objection at all to people refactoring and retrofitting for memory safety
426
00:27:53,400 --> 00:28:00,120
existing code bases that they cannot, for reasons of compatibility, resource constraint,
427
00:28:00,120 --> 00:28:01,680
others convert to Rust.
428
00:28:01,680 --> 00:28:08,280
I think we have to be very careful conflating expert level available, or tools available
429
00:28:08,280 --> 00:28:13,600
to expert level C++ developers from something like Rust that is just secure out of the box.
430
00:28:13,600 --> 00:28:17,760
The big thing that we've learned after 30 years of chasing memory safety is just because
431
00:28:17,760 --> 00:28:21,960
there is a way to do things safely doesn't mean people will.
432
00:28:21,960 --> 00:28:27,720
And I think the major, major, major advantage of Rust, Golang, Swift, and other languages
433
00:28:27,720 --> 00:28:30,600
is it's difficult to get wrong.
434
00:28:30,600 --> 00:28:35,480
And I would say that C++ remains inverted, which is it's difficult to get right.
435
00:28:35,480 --> 00:28:40,800
And as long as that's the case, I don't think we should recommend from a pure security perspective,
436
00:28:40,800 --> 00:28:43,400
people writing net new C++.
437
00:28:43,400 --> 00:28:47,760
That doesn't mean I would never discourage someone from improving security.
438
00:28:47,760 --> 00:28:52,960
For example, Herb Sutter and other folks who are awesome in the C++ community for driving
439
00:28:52,960 --> 00:28:59,560
in safety have invented things like GSL span, which does a lot for spatial safety and provides
440
00:28:59,560 --> 00:29:02,760
a container format that is spatially safe.
441
00:29:02,760 --> 00:29:08,800
But then you have things like temporal issues where smart pointers, garbage collectors,
442
00:29:08,800 --> 00:29:09,800
et cetera.
443
00:29:09,800 --> 00:29:14,440
And still, I don't know if there's a perfect standard.
444
00:29:14,440 --> 00:29:16,600
Google has things like Miracle Pointer, et cetera.
445
00:29:16,600 --> 00:29:18,720
But they aren't in standard part of the language.
446
00:29:18,720 --> 00:29:20,160
They aren't standard out of the box.
447
00:29:20,160 --> 00:29:25,400
And so until that's the case, I don't think it's quite comparable as a North Star to Rust.
448
00:29:25,400 --> 00:29:27,240
And I think that's an important point.
449
00:29:27,240 --> 00:29:30,160
The word that you mentioned there is North Star.
450
00:29:30,160 --> 00:29:33,160
And yeah, I do a good look.
451
00:29:33,160 --> 00:29:39,600
I've been developing in Windows, sorry, in C++ forever.
452
00:29:39,600 --> 00:29:40,600
And I love the language.
453
00:29:40,600 --> 00:29:41,600
I actually really do.
454
00:29:41,600 --> 00:29:44,400
And I actually really enjoy using modern C++.
455
00:29:44,400 --> 00:29:48,600
But to be honest with you, when it comes to writing new stuff, my preference is Rust.
456
00:29:48,600 --> 00:29:50,160
The tool chain is good.
457
00:29:50,160 --> 00:29:56,100
The compiler actually emits usable error messages that actually help you.
458
00:29:56,100 --> 00:30:02,360
So yeah, I'm a big fan, even though I brought up on C and C++ for a long, long time.
459
00:30:02,360 --> 00:30:06,040
And I think that's especially important in team environments, Michael.
460
00:30:06,040 --> 00:30:10,200
We know that you're one of the ranking experts in this, but I'm sure you could inform us
461
00:30:10,200 --> 00:30:14,200
like how many projects at Microsoft have you worked on where you're going to be the forever
462
00:30:14,200 --> 00:30:19,540
maintainer or isn't worked on by a team of folks with a variety of experiences?
463
00:30:19,540 --> 00:30:24,440
And so as long as you have to be an expert and know how to use these language tools to
464
00:30:24,440 --> 00:30:29,720
be safe, I think it's really hard to reach our goal, which is ultimately memory safety
465
00:30:29,720 --> 00:30:34,660
and complete eradication of certain classes of memory safety issues.
466
00:30:34,660 --> 00:30:38,480
So you brought up before that Mark Rusinovich said that new projects should be written in
467
00:30:38,480 --> 00:30:39,480
memory safe
languages.
468
00:30:39,480 --> 00:30:41,320
I assume that also means C sharp, right?
469
00:30:41,320 --> 00:30:42,320
Where it makes sense.
470
00:30:42,320 --> 00:30:43,400
Yeah, totally.
471
00:30:43,400 --> 00:30:49,760
I think there's a lot of focus on Rust because Rust has been sort of replacing C and C++
472
00:30:49,760 --> 00:30:51,920
in cases as a systems language.
473
00:30:51,920 --> 00:30:56,880
But let's be clear, like on platforms like Android that has a billion users, I mean,
474
00:30:56,880 --> 00:30:58,760
people have been writing Java.
475
00:30:58,760 --> 00:31:02,320
Most web-based services don't have a ton of memory safety issues.
476
00:31:02,320 --> 00:31:08,920
They suffer from logic issues because they're built in things like Python, C sharp, Golang,
477
00:31:08,920 --> 00:31:09,920
et cetera.
478
00:31:09,920 --> 00:31:14,240
And so, you know, and even JavaScript is memory safe.
479
00:31:14,240 --> 00:31:18,200
And so I think if you were to take sort of the, I forget who doesn't maybe stack overflow
480
00:31:18,200 --> 00:31:23,160
is to take the top languages every year that are written or most popular languages, I think
481
00:31:23,160 --> 00:31:26,880
probably six or seven out of 10 are probably memory safe.
482
00:31:26,880 --> 00:31:30,800
So what Rust is really filling in is that systems level high performance language and
483
00:31:30,800 --> 00:31:31,800
providing memory safety.
484
00:31:31,800 --> 00:31:32,800
All right.
485
00:31:32,800 --> 00:31:37,800
Hey, Gladys, is there anything else you want to ask?
486
00:31:37,800 --> 00:31:40,200
I think we're kind of done.
487
00:31:40,200 --> 00:31:41,200
If not.
488
00:31:41,200 --> 00:31:43,760
There is one thing I wanted to add that I think we haven't touched on that would be
489
00:31:43,760 --> 00:31:50,960
pretty cool is, you know, we talked about how to retrofit C and C++ with, you know,
490
00:31:50,960 --> 00:31:56,420
safer pointers or, you know, spatially safe containers like GSL span.
491
00:31:56,420 --> 00:32:00,640
But there's actually a really set of exciting technologies that I think can make this quite
492
00:32:00,640 --> 00:32:02,600
a bit easier at the operating system level.
493
00:32:02,600 --> 00:32:04,440
And that's memory tagging.
494
00:32:04,440 --> 00:32:13,040
So ARM has a standard called MTE that's actually available widespread on Pixel 8 phones and
495
00:32:13,040 --> 00:32:17,160
Intel, AMD and others, I think are looking into these technologies.
496
00:32:17,160 --> 00:32:22,360
But in a nutshell, what these memory tagging technologies do is allow an allocator at the
497
00:32:22,360 --> 00:32:23,360
system level.
498
00:32:23,360 --> 00:32:30,920
So, effectively, malloc to provide to the CPU the ranges of memory when memory is allocated
499
00:32:30,920 --> 00:32:32,560
to a pointer.
500
00:32:32,560 --> 00:32:37,400
And then what is really exciting is depending on variety of schemes from a coding the pointer
501
00:32:37,400 --> 00:32:44,400
to having, you know, bitmaps, the CPU can actually go and check every memory access
502
00:32:44,400 --> 00:32:48,180
and make sure it's within the expected boundaries of the allocation.
503
00:32:48,180 --> 00:32:54,520
And that is almost like sprinkling, you know, software checks all throughout the operating
504
00:32:54,520 --> 00:32:55,520
system.
505
00:32:55,520 --> 00:32:57,340
But it comes at, you know, you don't have to remember to do it.
506
00:32:57,340 --> 00:32:59,920
And it's done in a very high performance manner.
507
00:32:59,920 --> 00:33:02,720
So we've actually published several papers on that.
508
00:33:02,720 --> 00:33:05,020
Again, I can link in the show notes at the Microsoft level.
509
00:33:05,020 --> 00:33:09,080
But there's also been, you know, Google Project Zero and many folks have looked at this.
510
00:33:09,080 --> 00:33:13,360
But what you find, depending on the scheme, is that we get pretty close to the spatial
511
00:33:13,360 --> 00:33:19,480
safety of retrofitting C and C++ code without doing anything other than enlightening the
512
00:33:19,480 --> 00:33:20,920
allocator.
513
00:33:20,920 --> 00:33:24,400
And so I'm a big fan of this, especially as we're going through this transition to the
514
00:33:24,400 --> 00:33:29,240
Rust North Star of not only retrofitting your own code, but then code bases where we, you
515
00:33:29,240 --> 00:33:34,240
know, lost the source code or just, you know, don't have the experts to retrofit them.
516
00:33:34,240 --> 00:33:39,120
We can actually at least provide spatial safety with these architectures.
517
00:33:39,120 --> 00:33:44,680
And maybe even a step further is there is an exciting architecture based on ARM.
518
00:33:44,680 --> 00:33:50,920
And in fact, ARM shipped some experimental devices called Morello, which include what's
519
00:33:50,920 --> 00:33:55,120
called Cherry or a capabilities hardware language.
520
00:33:55,120 --> 00:33:59,120
And Cherry goes a step further than memory tagging in that every pointer actually has
521
00:33:59,120 --> 00:34:03,240
encoded into it things like, you know, what type of memory it is, whether it's code or
522
00:34:03,240 --> 00:34:06,320
data, whether it's a pointer, et cetera.
523
00:34:06,320 --> 00:34:10,600
And those capabilities that can be checked on every memory access.
524
00:34:10,600 --> 00:34:18,200
And so Cherry gives you a lot of primitives for building a memory safe operating system,
525
00:34:18,200 --> 00:34:20,560
regardless of the language you build in.
526
00:34:20,560 --> 00:34:27,960
And so Microsoft has even worked directly with the RISC-V community to build an ISA
527
00:34:27,960 --> 00:34:33,160
or instruction set that provides Cherry in RISC-V.
528
00:34:33,160 --> 00:34:38,960
So ARM's got Morello and has been, I think, promoting this and looking for a big technology
529
00:34:38,960 --> 00:34:44,020
partner to take on Cherry, which would provide more comprehensive
memory safety for existing
530
00:34:44,020 --> 00:34:45,540
operating systems.
531
00:34:45,540 --> 00:34:52,120
And then I know the RISC-V and Cherry foundations, Low Risk is actually the name of the organization,
532
00:34:52,120 --> 00:34:56,920
has been providing a bunch of, you know, basic developer boards and ships that people can
533
00:34:56,920 --> 00:34:58,320
go and hack on.
534
00:34:58,320 --> 00:35:02,280
So if you're interested in memory safety at all, beyond Rust, I think looking at these
535
00:35:02,280 --> 00:35:06,480
architectures as another tool in the toolbox to get to that North Star is really, really
536
00:35:06,480 --> 00:35:07,560
important.
537
00:35:07,560 --> 00:35:11,800
And I expect those to be hugely important as we transition towards a Rust-based or a
538
00:35:11,800 --> 00:35:12,800
memory safe feature.
539
00:35:12,800 --> 00:35:13,800
All right.
540
00:35:13,800 --> 00:35:16,800
Is there anything else you want to add Gladys?
541
00:35:16,800 --> 00:35:20,480
Because if not, I'll just move to the, you know, if you have one final thought.
542
00:35:20,480 --> 00:35:21,480
Okay.
543
00:35:21,480 --> 00:35:22,480
All right.
544
00:35:22,480 --> 00:35:23,480
All right.
545
00:35:23,480 --> 00:35:25,880
So David, you know about this, right?
546
00:35:25,880 --> 00:35:28,880
We asked for a final thought, just a simple single thought.
547
00:35:28,880 --> 00:35:30,440
You're okay with that?
548
00:35:30,440 --> 00:35:31,440
Sure.
549
00:35:31,440 --> 00:35:32,440
All right.
550
00:35:32,440 --> 00:35:37,360
So one thing we always ask our guests is if you had one final thought to leave our listeners
551
00:35:37,360 --> 00:35:40,200
with, what would it be?
552
00:35:40,200 --> 00:35:46,400
Yeah, I think thinking about how software is now running the base infrastructure for
553
00:35:46,400 --> 00:35:52,440
the world from everything from healthcare to food to industry is really important.
554
00:35:52,440 --> 00:35:57,040
The decisions we make as developers will be around much longer than probably any of us
555
00:35:57,040 --> 00:35:59,560
are alive and have larger consequences than we think.
556
00:35:59,560 --> 00:36:05,120
So thinking safety first and thinking like a car engineer does in the software space
557
00:36:05,120 --> 00:36:07,440
is something I really like encourage everyone to do.
558
00:36:07,440 --> 00:36:08,440
All right.
559
00:36:08,440 --> 00:36:11,160
Well, thank you so much for joining us this week.
560
00:36:11,160 --> 00:36:14,840
I'm always excited to nerd out about programming languages a little bit.
561
00:36:14,840 --> 00:36:15,840
Absolutely.
562
00:36:15,840 --> 00:36:16,840
It's always my thing.
563
00:36:16,840 --> 00:36:18,560
But yeah, again, thank you so much for joining us.
564
00:36:18,560 --> 00:36:22,400
I know you're ridiculously busy, especially with just everything that's going on with
565
00:36:22,400 --> 00:36:23,400
SFI.
566
00:36:23,400 --> 00:36:26,520
You can't keep me away from a memory safety conversation.
567
00:36:26,520 --> 00:36:27,520
Yeah, I know.
568
00:36:27,520 --> 00:36:28,520
I know.
569
00:36:28,520 --> 00:36:32,800
Yeah, we didn't even talk about analysis stuff, but we can do that for another day.
570
00:36:32,800 --> 00:36:33,800
We could get further.
571
00:36:33,800 --> 00:36:34,800
Yeah, let's do it again.
572
00:36:34,800 --> 00:36:35,800
All right.
573
00:36:35,800 --> 00:36:39,280
And yeah, so thanks again for joining us and to all our listeners out there.
574
00:36:39,280 --> 00:36:43,560
We hope you found this episode useful, nerdy, but hopefully useful.
575
00:36:43,560 --> 00:36:49,520
If you have any questions, you can just post it on our Twitter account and we'll happily
576
00:36:49,520 --> 00:36:50,600
get back to you.
577
00:36:50,600 --> 00:36:53,120
So stay safe and we'll see you next time.
578
00:36:53,120 --> 00:36:56,120
Thanks for listening to the Azure Security Podcast.
579
00:36:56,120 --> 00:37:02,960
You can find show notes and other resources at our website, azsecuritypodcast.net.
580
00:37:02,960 --> 00:37:07,800
If you have any questions, please find us on Twitter at Azure Setpod.
581
00:37:07,800 --> 00:37:31,680
Background music is from ccmixtor.com and licensed under the Creative Commons license.