TRANSKRİPT 🔗
1
00:00:01,126 --> 00:00:03,628
If you lose trust with your customer,
it's mostly
2
00:00:03,628 --> 00:00:07,507
crisis management, and trust
is the foundation of every business.
3
00:00:07,549 --> 00:00:12,512
It waits in the lower right
corner of your website, a small customer
4
00:00:12,512 --> 00:00:17,350
service chatbox with your company's logo
and a familiar phrase: How can we help?
5
00:00:17,851 --> 00:00:19,227
A customer types a question.
6
00:00:19,227 --> 00:00:23,148
Drift, the company behind the white label
chatbox remembers who they are,
7
00:00:23,148 --> 00:00:25,066
retrieves an earlier conversation
8
00:00:25,066 --> 00:00:28,903
and connects the exchange to the company's
records and Salesforce.
9
00:00:28,945 --> 00:00:30,780
The answer arrives, the window
10
00:00:30,780 --> 00:00:34,159
closes, and nothing about the moment
seems important.
11
00:00:34,534 --> 00:00:37,662
Across thousands of companies,
Drift handles
12
00:00:37,662 --> 00:00:40,790
tens of millions of conversations
every year.
13
00:00:41,124 --> 00:00:43,710
The chatbox is merely
the part people can see.
14
00:00:43,710 --> 00:00:47,964
Behind it sits something
far more powerful: Permission.
15
00:00:48,131 --> 00:00:51,259
Permission to recognize customers,
to read and update records,
16
00:00:51,259 --> 00:00:56,389
to connect conversations with email
and sales histories and support cases.
17
00:00:56,389 --> 00:01:00,310
Companies grant Drift that authority
deliberately, of course, it is
18
00:01:00,310 --> 00:01:04,606
what makes the platform
useful, powerful and profitable.
19
00:01:04,606 --> 00:01:10,612
And in the summer of 2025, it is
what made the platform so very dangerous.
20
00:01:10,779 --> 00:01:12,864
August 7th inside Salesloft,
21
00:01:12,864 --> 00:01:16,576
Drift’s parent company,
employees are sharing good news.
22
00:01:16,618 --> 00:01:19,788
Salesloft
and Clari have announced plans to merge.
23
00:01:19,871 --> 00:01:23,917
Together, they intend to create
what they call a new revenue
24
00:01:23,958 --> 00:01:27,170
AI powerhouse,
bringing customer conversations, buyer
25
00:01:27,170 --> 00:01:31,591
signals and revenue intelligence
into one connected platform.
26
00:01:31,841 --> 00:01:36,054
There are congratulatory messages,
optimistic posts, executives
27
00:01:36,054 --> 00:01:39,849
speaking about everything
these new connections will make possible.
28
00:01:39,891 --> 00:01:44,521
But somewhere beyond the celebration,
someone has already discovered
29
00:01:44,521 --> 00:01:47,315
what those connections
make possible for them.
30
00:01:47,315 --> 00:01:47,857
The following
31
00:01:47,857 --> 00:01:50,860
morning, requests
begin entering the Salesforce environments
32
00:01:50,860 --> 00:01:53,655
of companies using Drift.
They carry no malware.
33
00:01:53,655 --> 00:01:56,116
They exploit
no vulnerability in Salesforce.
34
00:01:56,116 --> 00:02:00,662
They do not come from some unknown
or suspicious application.
35
00:02:00,787 --> 00:02:03,081
They come through Drift.
36
00:02:03,081 --> 00:02:04,457
The application is trusted.
37
00:02:04,457 --> 00:02:08,711
The permissions are very real
and to the systems receiving the requests,
38
00:02:08,753 --> 00:02:10,046
nothing appears to be wrong.
39
00:02:10,046 --> 00:02:14,467
Of course, Drift is simply doing
what these companies authorized it to do.
40
00:02:14,843 --> 00:02:19,222
Except Drift is no longer
the only one deciding what it does next.
41
00:02:19,389 --> 00:02:22,392
Over the days that follow, information
begins leaving
42
00:02:22,392 --> 00:02:25,395
Salesforce environments around the world.
43
00:02:25,436 --> 00:02:29,357
Some of it belongs to technology
companies, some to cybersecurity firms,
44
00:02:29,357 --> 00:02:34,112
and some of it has been sitting inside
closed support cases for years.
45
00:02:34,154 --> 00:02:38,825
Names, conversations, technical details
and secrets that were never meant
46
00:02:38,825 --> 00:02:40,535
to become permanent records.
47
00:02:40,535 --> 00:02:44,122
By the time the first warnings
reach many of these organizations,
48
00:02:44,122 --> 00:02:45,081
the requests have stopped.
49
00:02:45,081 --> 00:02:48,626
The data is already gone,
and the most important part of the crime
50
00:02:48,626 --> 00:02:52,463
happened months before anyone knew
there was a crime at all.
51
00:02:52,463 --> 00:02:54,299
So who was acting through Drift?
52
00:02:54,299 --> 00:02:57,343
How did they acquire authority
that hundreds of companies
53
00:02:57,343 --> 00:02:58,970
had granted to someone else?
54
00:02:58,970 --> 00:03:02,390
And when every request appears legitimate.
55
00:03:02,432 --> 00:03:05,351
How do you recognize the intruder
behind it?
56
00:03:05,351 --> 00:03:09,189
This is the story of the Salesloft
Drift breach.
57
00:03:09,439 --> 00:03:13,568
To help us with this investigation,
we're joined by Volker Rath Field,
58
00:03:13,610 --> 00:03:17,238
CTO for Asia Pacific at Cloudflare.
59
00:03:17,405 --> 00:03:20,408
Cloudflare powers
20% of the internet and blocks
60
00:03:20,408 --> 00:03:24,454
more than 200 billion
cyber threats every day.
61
00:03:24,621 --> 00:03:27,582
Through its Global Connectivity Cloud, it
helps organizations connect
62
00:03:27,582 --> 00:03:31,169
and protect their applications,
employees, and networks.
63
00:03:31,169 --> 00:03:35,298
And we're very proud to have Cloudflare
sponsor today's episode.
64
00:03:35,298 --> 00:03:37,091
We'd like to thank them for their support.
65
00:03:37,091 --> 00:03:41,512
Before joining Cloudflare,
Volker spent more than 25 years
66
00:03:41,512 --> 00:03:42,513
helping some of the world's
67
00:03:42,513 --> 00:03:46,851
largest organizations navigate security,
cloud risk, and regulatory
68
00:03:46,851 --> 00:03:51,147
compliance through senior roles
at Amazon Web Services. BT,
69
00:03:52,315 --> 00:03:54,943
PWC,
Load full transcript
Hewlett-Packard, and Symantec.
70
00:03:54,943 --> 00:03:59,822
Today, he works with executive teams,
confronting exactly the kind of challenge
71
00:03:59,906 --> 00:04:04,619
at the center of this investigation:
What happens when trusted cloud
72
00:04:04,619 --> 00:04:08,331
access is turned against the organization
that granted it?
73
00:04:08,414 --> 00:04:11,417
Volker.
Welcome to The CISO Signal. Jeremy.
74
00:04:11,417 --> 00:04:13,461
Thank you for having me. It's
great to have you here.
75
00:04:13,461 --> 00:04:15,463
Are you ready to begin the investigation?
76
00:04:15,463 --> 00:04:16,506
Let's go into it.
77
00:04:16,506 --> 00:04:18,383
Let's get into it then.
78
00:04:18,383 --> 00:04:21,386
We are in the midst of a ceaseless war,
79
00:04:21,719 --> 00:04:24,889
not of bombs or bullets, but of breaches,
80
00:04:24,931 --> 00:04:28,268
firewalls and silent incursions.
81
00:04:28,601 --> 00:04:29,227
The targets?
82
00:04:29,227 --> 00:04:33,106
Our borders, our banks, our commerce,
and the critical
83
00:04:33,106 --> 00:04:37,151
infrastructure
that underpins a free civilization.
84
00:04:37,193 --> 00:04:40,405
The enemy is cloaked in code, fueled by
85
00:04:40,405 --> 00:04:43,866
greed, glory, and a desire for chaos.
86
00:04:43,908 --> 00:04:48,454
This is the story of the unseen
protectors, the nameless generals,
87
00:04:48,496 --> 00:04:52,583
the CISOs, chief information
security officers.
88
00:04:52,625 --> 00:04:56,963
They are the guardians at the gate,
the watchers on the wall,
89
00:04:57,130 --> 00:05:02,677
ever vigilant
and always listening for The CISO Signal.
90
00:05:04,846 --> 00:05:08,808
Act I: The Door Nobody Watched.
91
00:05:09,183 --> 00:05:13,021
When you're talking with CISOs
trying to help their business move faster
92
00:05:13,021 --> 00:05:16,858
without creating unnecessary risk,
how do you help them decide
93
00:05:16,858 --> 00:05:19,819
which applications
deserve that level of trust,
94
00:05:19,819 --> 00:05:23,406
and how much authority
those applications should really have?
95
00:05:23,990 --> 00:05:26,492
Interesting question,
because I had this conversation
96
00:05:26,492 --> 00:05:30,580
literally two hours ago with a CISO
and we talked about a particular
97
00:05:30,580 --> 00:05:32,457
application
the organization wanted to launch.
98
00:05:32,457 --> 00:05:33,833
And like usual,
99
00:05:33,833 --> 00:05:38,046
these kind of organizations end up very,
very quickly in ‘what do I need to buy?’
100
00:05:38,087 --> 00:05:42,425
‘What are the controls?’ Typically
over-indexing on technical controls,
101
00:05:42,425 --> 00:05:45,511
which are, by the way,
only 30% of the NIST or ISO framework.
102
00:05:45,553 --> 00:05:47,889
So security
is typically a technical problem isn't it.
103
00:05:47,889 --> 00:05:50,516
And then over-indexing
on protective controls.
104
00:05:50,516 --> 00:05:54,937
And when I hear these kind of things
‘what can I buy?’ ‘What should I buy?’
105
00:05:54,979 --> 00:05:59,025
Especially now with AI, AI firewall,
AI gateway, secure
106
00:05:59,067 --> 00:06:00,610
MCP servers, all these kind of things.
107
00:06:00,610 --> 00:06:03,237
I typically say,
let's step back for a second
108
00:06:03,237 --> 00:06:05,031
and think about what
we are trying to achieve.
109
00:06:05,031 --> 00:06:07,241
And what gets us closer to the answer
is actually not
110
00:06:07,241 --> 00:06:08,659
cyber security, it’s risk management.
111
00:06:08,659 --> 00:06:10,953
You don't spend $100 to protect ten.
112
00:06:10,953 --> 00:06:13,706
You don't spend $10 to protect a million.
113
00:06:13,706 --> 00:06:15,750
I ask the question,
what are we trying to protect here?
114
00:06:15,750 --> 00:06:19,253
In the first case,
it was an AI application voice-to-voice,
115
00:06:19,253 --> 00:06:23,007
all the data that was available to
this chat was public data.
116
00:06:23,132 --> 00:06:26,969
A customer types one last message:
Thanks, that fixed it.
117
00:06:26,969 --> 00:06:30,681
On the other side of the screen,
an employee clicks ‘Resolved’.
118
00:06:30,723 --> 00:06:32,266
The notification disappears,
119
00:06:32,266 --> 00:06:36,270
the conversation moves into the archive,
and both people carry on with their day.
120
00:06:36,312 --> 00:06:41,025
But behind the chat window,
nothing is disconnected.
121
00:06:41,109 --> 00:06:44,821
Drift’s access to the companies’
other systems remains active,
122
00:06:44,821 --> 00:06:49,367
waiting for the next conversation,
the next customer, the next update.
123
00:06:49,659 --> 00:06:51,869
The people have left.
124
00:06:51,869 --> 00:06:53,621
The connection remains.
125
00:06:53,621 --> 00:06:55,164
We talked about security and compliance.
126
00:06:55,164 --> 00:06:57,708
I said: Why are you talking about
security and compliance?
127
00:06:57,708 --> 00:07:00,169
The risk is very low.
There is only public data.
128
00:07:00,169 --> 00:07:01,921
There is no such thing like a breach.
129
00:07:01,921 --> 00:07:03,673
You can't steal public data.
130
00:07:03,673 --> 00:07:05,925
And there's another application
that handles
131
00:07:05,925 --> 00:07:08,219
sensitive information and also had AI
132
00:07:08,219 --> 00:07:10,513
and I said the conversation here
is entirely different.
133
00:07:10,513 --> 00:07:11,848
We have authenticated
134
00:07:11,848 --> 00:07:15,852
traffic going to sensitive data
and suddenly controls like principle
135
00:07:15,852 --> 00:07:19,981
of least privileges, segregation
of duties, Zero Trust, Defense in Depth,
136
00:07:19,981 --> 00:07:23,526
all the good things we have to offer in
security becomes much more important.
137
00:07:23,568 --> 00:07:24,986
Every interaction with Drift
138
00:07:24,986 --> 00:07:28,739
could quietly reach into Salesforce,
pulling customer records, updating
139
00:07:28,739 --> 00:07:33,661
opportunities, opening support cases
and syncing
information between systems.
140
00:07:33,786 --> 00:07:36,706
OAuth tokens worked silently
in the background, allowing
141
00:07:36,706 --> 00:07:41,836
trusted applications to authenticate
without asking anyone to log in again.
142
00:07:42,044 --> 00:07:46,382
Convenience had become infrastructure,
infrastructure had become trust,
143
00:07:46,382 --> 00:07:51,053
and trust is often worth
far more than a password.
144
00:07:51,095 --> 00:07:53,431
The person actually said, it was a CISO,
he said, look,
145
00:07:53,431 --> 00:07:56,934
I did a mistake here because I tried
to protect everything the same way.
146
00:07:56,934 --> 00:08:00,771
But somebody once said, if you try
to defend everything, you defend nothing.
147
00:08:00,813 --> 00:08:04,066
Nearly
all hacks are actually quite boring.
148
00:08:04,108 --> 00:08:08,237
They use the same techniques
and procedures and tactics.
149
00:08:08,279 --> 00:08:10,990
The top three haven't changed. It’s
vulnerabilities.
150
00:08:10,990 --> 00:08:15,786
It's over permissive access policy,
so IAM, and wrongly configured stuff.
151
00:08:15,786 --> 00:08:18,664
If we could remove all of these three,
we will be fine.
152
00:08:18,664 --> 00:08:21,000
Nothing crashed that morning.
Nothing disappeared.
153
00:08:21,000 --> 00:08:22,960
Nothing happened that anyone could see
154
00:08:22,960 --> 00:08:26,797
because the attackers
weren't trying to force their way in.
155
00:08:27,089 --> 00:08:29,800
According to investigations
later published by Google's
156
00:08:29,800 --> 00:08:33,679
Threat Intelligence Group
and incident disclosures from Salesloft,
157
00:08:33,721 --> 00:08:39,310
the attackers had found something
much more valuable...Someone else's keys.
158
00:08:39,519 --> 00:08:42,396
Quite often,
especially when we talk about SaaS-to-SaaS
159
00:08:42,396 --> 00:08:45,441
integrations, so the integrations
behind the integrations,
160
00:08:45,483 --> 00:08:48,444
you know, banks
stop talking about third-party risks.
161
00:08:48,444 --> 00:08:50,321
They use jargon like N-party risks.
162
00:08:50,321 --> 00:08:52,490
It's a dependency
on a dependency on a dependency.
163
00:08:52,490 --> 00:08:53,449
So that's the key thing.
164
00:08:53,449 --> 00:08:54,909
If you run IT,
165
00:08:54,909 --> 00:08:58,746
so from an operational perspective,
and if you work in cybersecurity,
166
00:08:58,788 --> 00:09:02,083
the two most important
things are visibility and control.
167
00:09:02,124 --> 00:09:03,834
Exactly how the attackers
168
00:09:03,834 --> 00:09:07,797
first compromised Salesloft
has never been publicly confirmed.
169
00:09:07,797 --> 00:09:11,217
Everything
after that could largely be reconstructed.
170
00:09:11,217 --> 00:09:15,304
First, the attackers gained access
to Salesloft’s GitHub account.
171
00:09:15,346 --> 00:09:19,642
They downloaded repositories,
added guest users and created workflows.
172
00:09:19,642 --> 00:09:23,145
From there, investigators
believe they found a path into Drift’s
173
00:09:23,145 --> 00:09:26,691
cloud infrastructure,
and inside that environment sat
174
00:09:26,774 --> 00:09:30,820
the delegated authority Drift carried
on behalf of hundreds of customers.
175
00:09:30,820 --> 00:09:34,407
What is known
is that OAuth credentials associated
176
00:09:34,407 --> 00:09:37,410
with Drift’s
Salesforce integration were compromised.
177
00:09:37,410 --> 00:09:40,454
They were just strings of characters,
of course, unremarkable
178
00:09:40,454 --> 00:09:42,039
to anyone glancing at them.
179
00:09:42,039 --> 00:09:47,378
But those strings represented identity,
not the identity of people,
180
00:09:47,461 --> 00:09:49,130
the identity of software
181
00:09:49,130 --> 00:09:53,551
that hundreds of organizations
had already chosen to trust.
182
00:09:53,551 --> 00:09:57,305
And if we authorize to two parties
183
00:09:57,346 --> 00:10:01,309
to interact with each other
and my authorization is required once,
184
00:10:01,350 --> 00:10:05,146
that for me is always a red flag because
Zero Trust says no permanent control.
185
00:10:05,146 --> 00:10:07,356
Trust placed comma every session.
186
00:10:07,356 --> 00:10:10,568
It's actually much worse than that because
we're not doing this just like that.
187
00:10:10,610 --> 00:10:14,113
We are calling our risk teams and say,
hey, can you do a risk assessment?
188
00:10:14,196 --> 00:10:17,408
Imagine walking into a secure office tower
189
00:10:17,450 --> 00:10:21,370
carrying the building manager's
all access badge.
190
00:10:21,662 --> 00:10:25,958
There's no need to break windows,
no need to pick locks, passageways
191
00:10:25,958 --> 00:10:32,173
simply open because the badge is genuine,
even though the person holding it is not.
192
00:10:32,214 --> 00:10:34,175
And the compliance teams
and they're looking into this
193
00:10:34,175 --> 00:10:38,596
and they're looking at their website
SOC2 click, ISO click, GDPR click.
194
00:10:38,638 --> 00:10:40,056
You do your research and say yeah
195
00:10:40,056 --> 00:10:43,726
the company is fine, ticks all the boxes
and they're compliant.
196
00:10:43,726 --> 00:10:45,853
And this is a one time check.
197
00:10:45,853 --> 00:10:49,106
After that contract is signed;
We authorize the integration.
198
00:10:49,106 --> 00:10:52,568
So we have a one time check and a one time
authorization.
199
00:10:52,568 --> 00:10:53,944
By early August.
200
00:10:53,944 --> 00:10:57,406
Investigators believe those credentials
were already being used.
201
00:10:57,448 --> 00:11:01,786
Google later identified activity
beginning as early as August 8th.
202
00:11:01,869 --> 00:11:05,498
From the outside,
nothing appeared unusual, and somewhere
203
00:11:05,498 --> 00:11:08,709
a support engineer was likely closing
a ticket before heading home.
204
00:11:08,709 --> 00:11:11,796
While five minutes later
an attacker might be opening
205
00:11:11,796 -->
00:11:14,840
that same case
from the other side of the world.
206
00:11:15,216 --> 00:11:19,178
Neither of them knew
they were looking at the same record.
207
00:11:19,220 --> 00:11:20,680
What can possibly go wrong?
208
00:11:20,680 --> 00:11:26,227
Especially when I then hand over all of
this responsibility to SaaS providers.
209
00:11:26,227 --> 00:11:28,979
But I don't have visibility
and I don't even have control.
210
00:11:28,979 --> 00:11:31,315
I don't know if they're really good.
211
00:11:31,315 --> 00:11:32,942
I don't know
if they're doing the wrong thing.
212
00:11:32,942 --> 00:11:34,151
They pass an audit.
213
00:11:34,151 --> 00:11:36,904
But compliance is for people
who don't know better.
214
00:11:36,904 --> 00:11:39,240
That lack of visibility and control...
215
00:11:39,240 --> 00:11:42,535
I always called it the SaaS cancer
and we haven’t cured it yet.
216
00:11:42,535 --> 00:11:47,123
Investigators would later trace suspicious
activity all the way back to March.
217
00:11:47,289 --> 00:11:50,668
The August breach wasn't really
the beginning, it was simply the moment
218
00:11:50,668 --> 00:11:54,630
the attackers chose to cash
in on months of preparation.
219
00:11:54,630 --> 00:11:57,466
Long before anyone noticed data
leaving Salesforce,
220
00:11:57,466 --> 00:12:01,303
someone had already been moving
quietly through Salesloft’s development
221
00:12:01,303 --> 00:12:05,766
environment, downloading code, creating
workflows and adding accounts
222
00:12:05,766 --> 00:12:09,019
while they map the territory
before the real theft ever began.
223
00:12:09,186 --> 00:12:13,858
The attackers then discovered something
even more valuable than customer records:
224
00:12:13,983 --> 00:12:17,778
support conversations
written at moments of vulnerability.
225
00:12:17,778 --> 00:12:22,616
An engineer may have copied a log, pasted
a configuration, or shared a temporary
226
00:12:22,658 --> 00:12:26,996
key, believing it would only ever be read
by the company they trusted.
227
00:12:27,288 --> 00:12:31,083
Months later, investigators discovered
the attackers were searching
228
00:12:31,083 --> 00:12:35,838
those same conversations for the next door
they could open.
229
00:12:35,880 --> 00:12:38,591
I don't understand,
you just said Zero Trust.
230
00:12:38,591 --> 00:12:40,634
How are we allowing all of that to happen?
231
00:12:40,634 --> 00:12:43,596
I have my love-hate relationship
with the term Zero Trust,
232
00:12:43,596 --> 00:12:47,057
because Zero Trust is the sharpest weapon
we have in security.
233
00:12:47,057 --> 00:12:50,186
And AI did not break Zero Trust so far.
234
00:12:50,311 --> 00:12:53,439
But if you follow the Zero Trust
principle, when I talk about Zero Trust
235
00:12:53,481 --> 00:12:55,775
and I need to explain this
because I work for a company
236
00:12:55,775 --> 00:12:59,278
who sells products with Zero Trust
and I'm not here to sell at all.
237
00:12:59,278 --> 00:13:03,115
When I talk about Zero Trust,
I talk about the NIST paper from 2010.
238
00:13:03,157 --> 00:13:06,118
Still a wonderful, wonderful paper.
239
00:13:06,118 --> 00:13:10,164
And if you take this really to your heart,
like assume state of compromise, micro
240
00:13:10,164 --> 00:13:13,918
segmentation, no long-lived
authentication, all these kind of things.
241
00:13:13,959 --> 00:13:16,837
It's not a framework,
it's guiding principles.
242
00:13:16,837 --> 00:13:19,840
And if you follow these guiding
principles, you'll be fine.
243
00:13:19,882 --> 00:13:25,721
The reality is, unfortunately,
that it became the most bastardized term.
244
00:13:25,763 --> 00:13:28,349
It became more
a marketing term for to sell.
245
00:13:28,349 --> 00:13:31,560
And in the year 2026,
246
00:13:31,685 --> 00:13:34,772
most of the stuff
is not operating like Zero Trust.
247
00:13:34,939 --> 00:13:40,236
The majority of enterprise workloads in
the cloud is old stuff,
248
00:13:40,236 --> 00:13:43,113
and we build new stuff,
and we struggle enough
249
00:13:43,113 --> 00:13:45,407
to actually follow Zero
Trust with the new stuff we built.
250
00:13:45,407 --> 00:13:47,660
That's an entirely different conversation,
but the reality is
251
00:13:47,660 --> 00:13:50,496
that the majority is the risk
we inherited.
252
00:13:50,496 --> 00:13:52,832
Individually,
none of it seemed remarkable.
253
00:13:52,832 --> 00:13:55,876
But together,
it formed something investigators
254
00:13:55,876 --> 00:13:59,129
believe the attackers had been searching
for all along.
255
00:13:59,171 --> 00:14:03,008
According to Google Threat Intelligence
Group, the campaign appeared focused
256
00:14:03,008 --> 00:14:06,136
on harvesting credentials
and other authentication material
257
00:14:06,178 --> 00:14:09,473
hidden inside
ordinary business conversations.
258
00:14:09,515 --> 00:14:13,269
The support platform
had quietly become a vault.
259
00:14:13,269 --> 00:14:19,066
And almost no one realized they had been
filling it with secrets for years.
260
00:14:19,066 --> 00:14:23,404
When I started my career, you know, what
were the two main controls in security?
261
00:14:23,445 --> 00:14:25,197
You know, firewall and antivirus.
262
00:14:25,197 --> 00:14:29,159
And today it's the three most important
263
00:14:29,493 --> 00:14:32,746
security controls are IAM, IAM, and IAM.
264
00:14:32,997 --> 00:14:34,415
So things have changed.
265
00:14:34,415 --> 00:14:39,545
But the design, the architecture of old
stuff is a risk, is a kind of sobering up.
266
00:14:39,545 --> 00:14:42,756
But also I believe
we kind of only delayed it.
267
00:14:42,798 --> 00:14:45,092
And then another clue surfaced.
268
00:14:45,092 --> 00:14:49,138
Google reported that credentials
associated with Drift’s Google Workspace
269
00:14:49,138 --> 00:14:54,184
email integration were also abused against
a very small number of organizations
270
00:14:54,184 --> 00:14:55,561
using that feature.
271
00:14:55,561 --> 00:14:59,899
Importantly, Google found no evidence
that Google Workspace itself
272
00:14:59,899 --> 00:15:02,818
had been compromised.
AI is changing a lot.
273
00:15:02,818 --> 00:15:05,529
All this old stuff is not AI compliant.
274
00:15:05,529 --> 00:15:10,242
AI is dominating the future, I would say,
in terms of data governance, integration,
275
00:15:10,284 --> 00:15:15,164
API driven, all this kind of stuff,
this old stuff doesn't serve AI very well.
276
00:15:15,164 --> 00:15:18,208
That's why AI is still kind of hamstrung.
277
00:15:18,208 --> 00:15:21,337
The old stuff is still,
the data governance is terrible.
278
00:15:21,378 --> 00:15:24,048
The access,
the entitlement models is not...
279
00:15:24,048 --> 00:15:28,594
is very, very different
of what we need to adopt AI at scale.
280
00:15:28,636 --> 00:15:31,639
I believe there's a time coming
and is probably coming soon, where we go
281
00:15:31,639 --> 00:15:37,353
into a big
wave of of not migrate and re-platform,
282
00:15:38,020 --> 00:15:41,857
but to actually retire, rearchitect,
renew.
283
00:15:41,899 --> 00:15:46,654
This is why we need to continually talk
about the principle of Zero Trust.
284
00:15:46,654 --> 00:15:49,156
But we cannot afford
to make this mistake again.
285
00:15:49,156 --> 00:15:53,327
And what may be even more unsettling
is that inside
286
00:15:53,369 --> 00:15:56,914
thousands of perfectly ordinary
support conversations
287
00:15:56,914 --> 00:16:01,293
lie credentials
capable of opening countless doors.
288
00:16:01,585 --> 00:16:02,252
I just want
289
00:16:02,252 --> 00:16:06,256
to take a moment to thank everyone
who's already subscribed to the channel.
290
00:16:06,256 --> 00:16:09,510
And if you haven't yet, please
consider subscribing,
291
00:16:09,510 --> 00:16:12,513
like the video, and let us know
what you think in the comments.
292
00:16:12,513 --> 00:16:15,933
If you know another security leader
who would enjoy The CISO Signal,
293
00:16:15,975 --> 00:16:17,685
please share it with them.
294
00:16:17,685 --> 00:16:19,812
It really helps us grow the show.
295
00:16:19,812 --> 00:16:24,608
And now Act Two: What The Keys Unlocked.
296
00:16:24,650 --> 00:16:31,323
Investigators eventually
gave the intruders a designation UNC6395.
297
00:16:31,532 --> 00:16:34,076
Not a name, simply a label for a group
298
00:16:34,076 --> 00:16:37,955
whose movement could be reconstructed
more easily than its identity.
299
00:16:37,955 --> 00:16:41,000
And by the time anyone realized
something was wrong.
300
00:16:41,041 --> 00:16:45,421
The group was no longer looking for a way
in but deciding what to take.
301
00:16:45,546 --> 00:16:51,510
According to reports, activity attributed
to UNC6395 unfolded methodically.
302
00:16:51,510 --> 00:16:54,763
The attackers behaved
like experienced surveyors walking through
303
00:16:54,763 --> 00:16:59,059
an unfamiliar building,
and Salesforce became their map.
304
00:16:59,226 --> 00:17:01,353
Every customer account,
every support case,
305
00:17:01,353 --> 00:17:05,315
every opportunity
represented another room to explore.
306
00:17:05,524 --> 00:17:08,193
But the real objective wasn't
Salesforce itself.
307
00:17:08,193 --> 00:17:11,196
It was what Salesforce revealed.
308
00:17:11,196 --> 00:17:15,826
Passwords, API keys, Amazon Web Service
credentials, Snowflake
309
00:17:15,826 --> 00:17:20,914
authentication tokens, and forgotten paths
into other systems.
310
00:17:20,956 --> 00:17:24,501
Individually,
they looked harmless, but together
311
00:17:24,501 --> 00:17:28,005
they were enough
to unlock the next breach.
312
00:17:28,088 --> 00:17:32,926
Every victim had become
a potential stepping stone to the next.
313
00:17:33,093 --> 00:17:37,556
The attack really exposed
the difference between identity and trust.
314
00:17:37,556 --> 00:17:42,061
An application can still be authorized
even after the trust behind
315
00:17:42,061 --> 00:17:43,395
it has been compromised.
316
00:17:43,395 --> 00:17:48,484
So beyond proving
who or what is making the request?
317
00:17:48,525 --> 00:17:49,568
What do you think
318
00:17:49,568 --> 00:17:50,235
should determine
319
00:17:50,235 --> 00:17:53,155
whether that request
deserves to be trusted?
320
00:17:53,155 --> 00:17:55,532
The question is kind of obsolete
if you don't
321
00:17:55,532 --> 00:17:57,242
operate on the basis of trust.
322
00:17:57,242 --> 00:17:58,827
The trust was established
323
00:17:58,827 --> 00:18:02,122
not within the session,
and that was ultimately the problem.
324
00:18:02,122 --> 00:18:05,250
So the trust was established
not in a computer,
325
00:18:05,250 --> 00:18:07,503
but in a PDF with two signatures.
326
00:18:07,503 --> 00:18:10,964
We need to think about what happened
was the attack started at GitHub.
327
00:18:10,964 --> 00:18:16,386
And the first slam dunk of the attacker
was when they found hard coded
328
00:18:16,386 --> 00:18:17,471
credentials.
329
00:18:17,471 --> 00:18:18,889
Google's investigators
330
00:18:18,889 --> 00:18:23,102
eventually understood what the attackers
had been searching for all along.
331
00:18:23,352 --> 00:18:25,312
The kind of information that accumulates
332
00:18:25,312 --> 00:18:29,316
naturally when engineers collaborate
to solve urgent problems.
333
00:18:29,316 --> 00:18:33,570
This wasn't an attacker manually wandering
from one company to the next,
334
00:18:33,570 --> 00:18:35,906
but a Python based automation
335
00:18:35,906 --> 00:18:39,618
repeating the same playbook across
hundreds of organizations.
336
00:18:39,785 --> 00:18:43,705
According to public reporting,
the campaign was disciplined enough
337
00:18:43,705 --> 00:18:48,585
that the attackers deleted Salesforce
query jobs after extricating data,
338
00:18:48,585 --> 00:18:52,339
an apparent attempt
to erase evidence of their searches.
339
00:18:52,339 --> 00:18:56,301
It wasn't perfect,
investigators
only knew they had been there because
340
00:18:56,343 --> 00:18:57,845
other logs survived.
341
00:18:57,845 --> 00:19:01,348
But it did reveal something important:
the attackers
342
00:19:01,348 --> 00:19:04,351
understood exactly what they had found.
343
00:19:04,726 --> 00:19:06,103
This is negligent.
344
00:19:06,103 --> 00:19:10,440
So we’re now, first of all, operating
not in a Zero Trust environment,
345
00:19:10,482 --> 00:19:14,695
then doing the wrong thing on top of it,
which increases the risk N times.
346
00:19:14,736 --> 00:19:18,782
And then they pivoted from there,
and then they ultimately got access
347
00:19:18,782 --> 00:19:21,910
to refresh
tokens used only by a part, you know,
348
00:19:21,994 --> 00:19:26,206
and then they got access
to all refresh tokens at a single place?!?
349
00:19:26,206 --> 00:19:29,418
That’s the equivalent of I'm
not only leaving my gold
350
00:19:29,418 --> 00:19:31,879
watch on the dinner table
and leave the door open,
351
00:19:31,879 --> 00:19:35,299
I take all my jewelry and put it on
the dinner table and leave the door open.
352
00:19:35,299 --> 00:19:37,342
And this is not uncommon
353
00:19:37,342 --> 00:19:41,597
because these applications...I
talked about the entitlement model.
354
00:19:41,597 --> 00:19:45,142
One company would later
provide the clearest public view
355
00:19:45,225 --> 00:19:48,228
of what happened next: Cloudflare.
356
00:19:48,228 --> 00:19:51,982
On August 9th, according to Cloudflare’s
own forensic timeline,
357
00:19:52,024 --> 00:19:55,861
the attackers began reconnaissance
against its Salesforce environment.
358
00:19:55,903 --> 00:19:59,615
Three days later, using compromised Drift
integration credentials,
359
00:19:59,615 --> 00:20:01,116
they gained access.
360
00:20:01,116 --> 00:20:03,785
No vulnerability in Salesforce
had been exploited.
361
00:20:03,785 --> 00:20:07,873
No Cloudflare employee
had typed a password into a fishing page.
362
00:20:08,040 --> 00:20:11,001
The attackers arrived carrying
trusted credentials
363
00:20:11,001 --> 00:20:15,631
issued by an application
every customer had already approved.
364
00:20:15,672 --> 00:20:19,009
For the next several days,
they quietly exported customer contact
365
00:20:19,009 --> 00:20:22,221
details, support
cases, technical conversations
366
00:20:22,221 --> 00:20:24,973
and when the first key they tried
didn't work,
367
00:20:24,973 --> 00:20:29,728
they simply went back into the stolen data
until they found one that did.
368
00:20:30,437 --> 00:20:34,191
It took three months to prepare,
three days to understand the environment,
369
00:20:34,191 --> 00:20:37,402
and about three minutes to steal the data.
370
00:20:37,486 --> 00:20:40,072
Here, we had the same thing.
There was an entitlement model.
371
00:20:40,072 --> 00:20:44,159
There is a contract with vendor A, vendor
B, vender C, and suddenly they said, oh,
372
00:20:44,201 --> 00:20:45,744
you know,
I
have this chat app...ah this chat app
373
00:20:46,828 --> 00:20:48,288
needs to have access to everything.
374
00:20:48,288 --> 00:20:49,831
Oh, there's another integration.
375
00:20:49,831 --> 00:20:52,084
There is the Snowflake integration.
376
00:20:52,084 --> 00:20:54,086
Where does that live? On Amazon.
377
00:20:54,086 --> 00:20:56,880
How does that actually integrate
into Salesforce?
378
00:20:56,880 --> 00:20:57,839
It uses keys.
379
00:20:57,839 --> 00:21:00,717
Where the keys? In Secrets Manager.
380
00:21:00,717 --> 00:21:03,845
But I have the key to Secrets Manager
because in all the refresh tokens
381
00:21:03,845 --> 00:21:04,888
there was one.
382
00:21:04,888 --> 00:21:07,099
Then the hacker
probably like in the real world,
383
00:21:07,099 --> 00:21:11,895
they had a bunch of keys and they’re like,
‘oh that works!’ Cloudflare later
384
00:21:11,937 --> 00:21:13,689
warned customers to treat anything
385
00:21:13,689 --> 00:21:17,693
submitted through those affected
support cases as potentially compromised,
386
00:21:17,693 --> 00:21:22,030
then became one of the investigations
most sobering discoveries.
387
00:21:22,197 --> 00:21:27,661
Cloudflare identified 104
API tokens inside the exposed data.
388
00:21:27,744 --> 00:21:30,747
Investigators found no evidence
those tokens had been abused,
389
00:21:30,747 --> 00:21:34,167
but the company rotated every one of them
anyway.
390
00:21:34,501 --> 00:21:37,754
104 potential open safes.
391
00:21:38,255 --> 00:21:42,426
104 opportunities
for the breach to spread.
392
00:21:42,592 --> 00:21:48,181
When you're talking with CISOs
today, has third-party risk become less
393
00:21:48,181 --> 00:21:52,144
about trusting vendors
and more about understanding just how much
394
00:21:52,269 --> 00:21:55,272
authority we've, we've delegated to them?
395
00:21:55,355 --> 00:21:58,608
Third-party risk is top of mind,
especially
396
00:21:59,401 --> 00:21:59,943
in industries
397
00:21:59,943 --> 00:22:03,655
that deal with very, very sensitive data,
like the highly regulated data,
398
00:22:03,780 --> 00:22:07,534
financial, personal, health
information, critical infrastructure.
399
00:22:07,576 --> 00:22:08,827
Third-party risk is top of mind
400
00:22:08,827 --> 00:22:10,787
because third-party risks
are becoming overwhelming.
401
00:22:10,787 --> 00:22:13,040
We have third-party
risks in multiple dimensions.
402
00:22:13,040 --> 00:22:16,209
So we have them on the software
and we have them in a service supply chain
403
00:22:16,209 --> 00:22:18,170
and that is constantly being compromised.
404
00:22:18,170 --> 00:22:21,423
Cloudflare also stated
there was no evidence that its production
405
00:22:21,423 --> 00:22:25,427
systems or customer services
had been compromised through the incident.
406
00:22:25,469 --> 00:22:28,930
The breach had reached valuable
information, but investigators found
407
00:22:28,972 --> 00:22:33,268
no indication that it had crossed into
the company's operational infrastructure.
408
00:22:33,435 --> 00:22:36,646
That distinction mattered
because it highlighted
409
00:22:36,646 --> 00:22:40,567
the uncomfortable reality
facing every affected organization.
410
00:22:40,692 --> 00:22:44,529
The real danger wasn't confined to
what had already been stolen.
411
00:22:44,571 --> 00:22:48,992
It was what that stolen information
might unlock next.
412
00:22:48,992 --> 00:22:50,494
So especially since
413
00:22:50,494 --> 00:22:55,665
AI and whatever we’ve seen also way
more GitHub hacks than ever before.
414
00:22:55,665 --> 00:22:58,210
So we talked about the services supply
chain.
415
00:22:58,210 --> 00:23:00,045
You're consuming services
from a third party.
416
00:23:00,045 --> 00:23:02,798
They use more and more
AI to provide you with services.
417
00:23:02,798 --> 00:23:05,384
And the supply chain is not linear.
418
00:23:05,384 --> 00:23:11,014
It's like a tree and it branches out
and it is a third and fourth layer.
419
00:23:11,098 --> 00:23:13,517
At first every breach looked isolated.
420
00:23:13,517 --> 00:23:15,977
Then another company
reported the same pattern.
421
00:23:15,977 --> 00:23:17,396
Then another.
422
00:23:17,396 --> 00:23:18,855
Investigators stopped
423
00:23:18,855 --> 00:23:22,109
looking at the victims and started
looking for what connected them.
424
00:23:22,234 --> 00:23:25,862
Different industries, different security
teams, different environments.
425
00:23:25,946 --> 00:23:29,699
The only thing they all shared was Drift.
426
00:23:30,075 --> 00:23:32,994
It wasn't hundreds of separate breaches.
427
00:23:32,994 --> 00:23:38,959
It was one compromise
echoing through hundreds of organizations.
428
00:23:38,959 --> 00:23:45,173
You have a myriad of interconnections
and dependencies, and therefore
429
00:23:45,215 --> 00:23:50,887
potential vulnerability, domino chains
that are waiting to fall, super important.
430
00:23:50,929 --> 00:23:55,350
My advice is we need to rethink cyber
security from the ground up.
431
00:23:55,350 --> 00:23:59,312
We need to challenge the status quo
because what we did yesterday
432
00:23:59,354 --> 00:24:00,313
doesn't work anymore.
433
00:24:00,313 --> 00:24:02,232
That doesn't mean
that we need to trash everything.
434
00:24:02,232 --> 00:24:05,277
But I believe security today is a car
435
00:24:05,277 --> 00:24:09,364
and AI security from the attacker
side is a jet engine.
436
00:24:09,406 --> 00:24:15,120
If we’re adding wider wheels,
tinted windows and a spoiler to the car,
437
00:24:15,120 --> 00:24:19,833
this is a continuous improvement,
we will never turn it into a jet engine.
438
00:24:19,833 --> 00:24:21,209
So we're dealing with
439
00:24:21,209 --> 00:24:25,005
a fundamental different threat landscape,
which is fueled by AI,
440
00:24:25,088 --> 00:24:29,259
which is coming like
post-quantum encryption, which is fueled
441
00:24:29,259 --> 00:24:32,762
by increased
activity of nation states,
which is...
442
00:24:32,762 --> 00:24:34,264
and I can go on and on.
443
00:24:34,264 --> 00:24:38,059
As Google's investigation expanded,
the victim list continued to grow.
444
00:24:38,185 --> 00:24:41,354
Public reporting and regulatory
guidance would eventually indicate
445
00:24:41,354 --> 00:24:45,442
that more than 700 organizations
were affected in some capacity.
446
00:24:45,442 --> 00:24:48,069
Although the level of exposure varied
and many companies
447
00:24:48,069 --> 00:24:51,490
never publicly disclosed
exactly what had been accessed.
448
00:24:51,531 --> 00:24:55,994
Some of those organizations
specialized in cybersecurity.
449
00:24:56,036 --> 00:24:59,164
Others invested heavily in mature
security programs.
450
00:24:59,164 --> 00:25:01,708
It made little difference in the end.
451
00:25:01,708 --> 00:25:05,420
The attackers weren't defeating
each company's defenses individually.
452
00:25:05,670 --> 00:25:09,716
They had compromised trust at its source.
453
00:25:09,758 --> 00:25:13,678
We are actually heavily innovating in
how we believe we can solve the Zero
454
00:25:13,678 --> 00:25:15,889
Trust problem in the SaaS ecosystem,
455
00:25:15,889 --> 00:25:18,850
we call it Saas-to-Saas Proxy
and we just say, okay,
456
00:25:18,850 --> 00:25:21,186
if we don't have Zero Trust
in, let's put it in.
457
00:25:21,186 --> 00:25:24,189
If I need to trust two parties
that work together
458
00:25:24,189 --> 00:25:27,192
and act on my behalf,
I maintain half of the key.
459
00:25:27,192 --> 00:25:29,402
We believe that we need to rebuild trust.
460
00:25:29,402 --> 00:25:34,324
That means follow the principles of Zero
Trust, implement visibility and control
461
00:25:34,324 --> 00:25:39,329
so that we do not have to wait anymore
until we see some anomalies somewhere
462
00:25:39,329 --> 00:25:44,334
else, like in the AWS environment,
if we're lucky and say,
463
00:25:44,376 --> 00:25:45,335
oh, what's happening here?
464
00:25:45,335 --> 00:25:49,130
And we have to go ten steps
back to actually understand, oh my God.
465
00:25:49,172 --> 00:25:52,592
By August 20th,
the response had finally caught up.
466
00:25:52,634 --> 00:25:56,846
Salesloft, working with Salesforce,
revoked the compromised OAuth
467
00:25:56,888 --> 00:25:59,849
access and refresh tokens
associated with Drift.
468
00:25:59,849 --> 00:26:02,394
Salesforce
removed the application from AppExchange.
469
00:26:02,394 --> 00:26:06,731
While investigators work to understand
the full scope of the compromise.
470
00:26:06,940 --> 00:26:12,904
The trusted door had finally been closed,
but every document was already copied,
471
00:26:12,904 --> 00:26:16,658
every credential already harvested,
every support conversation
472
00:26:16,658 --> 00:26:20,912
had already been downloaded,
had left the building with the attackers.
473
00:26:21,162 --> 00:26:23,915
Revoking
a key can stop someone coming back.
474
00:26:23,915
--> 00:26:28,962
But it cannot erase the vast library
of private conversations, harvested
475
00:26:28,962 --> 00:26:33,550
credentials and corporate secrets
the attackers had already escaped with.
476
00:26:35,385 --> 00:26:39,556
Act Three: What Trust Leaves Behind.
477
00:26:40,515 --> 00:26:43,935
The first signs of recovery
didn't look like victory.
478
00:26:43,935 --> 00:26:45,478
They looked like work.
479
00:26:45,478 --> 00:26:47,105
Across hundreds of organizations,
480
00:26:47,105 --> 00:26:51,026
security teams began the same
painstaking cleansing ritual.
481
00:26:51,026 --> 00:26:54,487
It was the kind of response
every incident commander knows well:
482
00:26:54,779 --> 00:26:57,657
quiet, methodical, and driven
483
00:26:57,657 --> 00:27:00,660
by a single, unsettling question.
484
00:27:00,869 --> 00:27:03,663
What else did they find?
485
00:27:03,663 --> 00:27:06,166
The concept of this half key and half key?
486
00:27:06,166 --> 00:27:08,877
And how does that work
practically in real terms?
487
00:27:08,877 --> 00:27:09,586
And what's the sort of
488
00:27:09,586 --> 00:27:13,423
the speed of that back and forth operation
of, of getting approval for that?
489
00:27:13,423 --> 00:27:14,799
Can you tell us a little bit
more about it?
490
00:27:14,799 --> 00:27:16,259
The answer is actually quite simple.
491
00:27:16,259 --> 00:27:19,888
We do the same thing
with, with, OAuth, with IAM, you know?
492
00:27:19,929 --> 00:27:21,222
You have an identity provider.
493
00:27:21,222 --> 00:27:22,682
You don't do it typically yourself.
494
00:27:22,682 --> 00:27:24,976
So you have this trusted
man in the middle.
495
00:27:24,976 --> 00:27:26,811
How do you connect to a website?
496
00:27:26,811 --> 00:27:29,814
Well, you use HTTPS. What is HTTPS?
497
00:27:29,814 --> 00:27:31,483
Well there's a certificate.
498
00:27:31,483 --> 00:27:34,152
Well how do certificates work? Well,
there's a trusted third party.
499
00:27:34,152 --> 00:27:36,571
I cannot give you a certificate
and say valid.
500
00:27:36,571 --> 00:27:39,949
There needs to be a trusted third party
that we all trust.
501
00:27:39,949 --> 00:27:45,163
And they say if it's valid
and you ask them, you don't ask me.
502
00:27:45,205 --> 00:27:51,461
The simple answer is we go from a two-way
trusted model to what we're doing
503
00:27:51,461 --> 00:27:55,173
for a long time, for logins and internet
and certificates, and everything...
504
00:27:55,632 --> 00:27:59,844
For many organizations,
the story was only beginning to unfold.
505
00:27:59,886 --> 00:28:02,931
Organizations were advised to treat
not only compromised
506
00:28:02,931 --> 00:28:05,100
Salesforce tokens as suspect,
but potentially
507
00:28:05,100 --> 00:28:09,104
any authentication material
connected to Drift’s integrations.
508
00:28:09,145 --> 00:28:12,816
Cloudflare later disclosed that
it disabled the affected Drift account,
509
00:28:12,857 --> 00:28:16,820
revoking credentials, removing Salesloft
software and browser extensions,
510
00:28:16,986 --> 00:28:20,031
suspending a third party
Salesforce integrations,
511
00:28:20,031 --> 00:28:23,034
and rotating
exposed authentication material.
512
00:28:23,284 --> 00:28:27,831
It wasn't an attempt to repair
what had happened - that was impossible.
513
00:28:27,956 --> 00:28:31,584
It was an attempt to prevent yesterday's
attack
514
00:28:31,626 --> 00:28:34,754
from becoming tomorrow's
even larger breach.
515
00:28:35,213 --> 00:28:37,757
When I was young
and the idea of eBay came up
516
00:28:37,757 --> 00:28:41,553
and I thought, why should I buy something
from people I don't know, not even trust?
517
00:28:41,594 --> 00:28:42,178
They’re selling me
518
00:28:42,178 --> 00:28:46,266
something that is used
and I cannot even touch it and inspect it.
519
00:28:46,266 --> 00:28:47,851
And I sent you money.
520
00:28:47,851 --> 00:28:49,102
I thought eBay was stupid,
well I was stupid.
521
00:28:49,102 --> 00:28:50,645
Well, I was, I must admit,
522
00:28:50,645 --> 00:28:54,858
after all, but the question
I think was valid initially, right?
523
00:28:54,899 --> 00:28:58,987
But why is eBay not stupid to use?
524
00:28:58,987 --> 00:29:03,324
Because eBay invented that three-way
trust system with money, PayPal.
525
00:29:03,324 --> 00:29:06,995
And they invented a three-way trust system
with hey, if something is wrong,
526
00:29:07,036 --> 00:29:09,622
you get your money back,
the assurance behind it,
527
00:29:09,622 --> 00:29:13,418
the confidence behind it,
and you can write them
528
00:29:13,418 --> 00:29:17,547
and we say, okay, is it a trusted
vendor, 99.97%, and feedback.
529
00:29:17,589 --> 00:29:21,676
And then suddenly we establish trust
and the rest is history.
530
00:29:21,718 --> 00:29:23,136
eBay did it a long time ago.
531
00:29:23,136 --> 00:29:26,264
We're doing it with every HTTPS connection
today.
532
00:29:26,264 --> 00:29:28,683
We do it with every cloud login today.
533
00:29:28,683 --> 00:29:31,770
But when it comes to SaaS-to-SaaS
integrations, we forget everything.
534
00:29:31,811 --> 00:29:34,022
We go back to the 1990s.
535
00:29:34,022 --> 00:29:37,901
But at Okta, one stolen token met
a different result.
536
00:29:37,984 --> 00:29:41,863
The credential was valid,
but Okta had restricted
537
00:29:41,863 --> 00:29:45,992
Drift’s access
to approved network locations, only.
538
00:29:45,992 --> 00:29:48,119
The attackers had the right key.
539
00:29:48,119 --> 00:29:51,122
They were just using it from the wrong
address.
540
00:29:51,372 --> 00:29:52,874
Access was denied.
541
00:29:52,874 --> 00:29:55,543
Elsewhere, possessing Drift’s
identity had been enough.
542
00:29:55,543 --> 00:30:00,924
But here the system asked
one more question - not only who are you,
543
00:30:00,924 --> 00:30:04,385
but should you be here?
544
00:30:04,844 --> 00:30:08,473
That's a question about incident response
best
practices
545
00:30:08,473 --> 00:30:10,475
and they only work
if you've done your homework.
546
00:30:10,475 --> 00:30:12,519
So have you established
visibility and control?
547
00:30:12,519 --> 00:30:15,480
I mean, I was involved
in a couple of investigations,
548
00:30:15,480 --> 00:30:18,358
and I always found it shocking
how little data we had to work with.
549
00:30:18,358 --> 00:30:20,860
So people say, oh,
we need logs for audits.
550
00:30:20,860 --> 00:30:24,614
I said, yeah,
but you also need logs for investigations.
551
00:30:24,614 --> 00:30:26,991
So the first question is,
what are we dealing with?
552
00:30:26,991 --> 00:30:29,160
You cannot fix the problem immediately.
553
00:30:29,160 --> 00:30:33,081
You need to make sure that you prevent
further things from happening.
554
00:30:33,081 --> 00:30:35,333
We need to remind ourselves that typically
555
00:30:35,333 --> 00:30:39,295
when we have 1 or 2 problems,
we will have ten problems very quickly.
556
00:30:39,337 --> 00:30:42,590
Where can we put that hand
between the dominoes?
557
00:30:42,799 --> 00:30:47,846
It's sometimes a hard conversation because
you cannot pull the plug on the bank.
558
00:30:47,887 --> 00:30:51,641
You cannot pull the plug
in an important governance service, health
559
00:30:51,641 --> 00:30:54,644
service; Extreme example,
somebody hacks the police.
560
00:30:54,853 --> 00:30:56,145
Are you shutting down the police?
561
00:30:56,145 --> 00:30:59,899
Typically
the answer is especially in the cloud,
562
00:31:00,108 --> 00:31:04,195
the three most important security
controls are IAM, IAM, and IAM.
563
00:31:04,237 --> 00:31:08,449
And they’re actually pretty good,
if you know what you're doing.
564
00:31:08,491 --> 00:31:10,326
We can immediately kill sessions.
565
00:31:10,326 --> 00:31:12,328
We can immediately activate accounts.
566
00:31:12,328 --> 00:31:16,708
Perhaps the most unsettling part of
the investigation is what remains unknown.
567
00:31:16,833 --> 00:31:20,545
Public reporting has not identified
the attackers with absolute certainty.
568
00:31:20,587 --> 00:31:25,049
The complete list of affected
organizations has never been released.
569
00:31:25,174 --> 00:31:29,387
No public accounting exists
for every credential that was harvested,
570
00:31:29,387 --> 00:31:32,515
every conversation that was copied,
or every downstream
571
00:31:32,515 --> 00:31:35,518
system
that those secrets may have unlocked.
572
00:31:35,685 --> 00:31:39,522
Investigators know where the attackers
wandered, but they cannot say
573
00:31:39,522 --> 00:31:44,736
with confidence what keys were taken
or which doors those keys might unlock.
574
00:31:44,736 --> 00:31:47,822
In cybercrime, that uncertainty
575
00:31:48,031 --> 00:31:50,825
has a very long shelf life.
576
00:31:50,825 --> 00:31:52,619
So we don't have to deactivate a service.
577
00:31:52,619 --> 00:31:55,288
We don't have to deactivate everything.
We just...
578
00:31:55,288 --> 00:31:59,626
we need enough information
to make the right decisions:
579
00:31:59,626 --> 00:32:02,587
What accounts and sessions
we need to kill.
580
00:32:02,629 --> 00:32:04,339
So a little bit more surgical.
581
00:32:04,339 --> 00:32:06,883
That of course requires some experience.
582
00:32:06,883 --> 00:32:11,429
So, you know, preparedness
in the right logs and the right training
583
00:32:11,429 --> 00:32:15,350
and the right incident response playbooks
and automation of incident response
584
00:32:15,350 --> 00:32:16,935
playbooks is becoming more important
585
00:32:16,935 --> 00:32:19,187
because everything now
happens at machine speed.
586
00:32:19,187 --> 00:32:20,855
I also would say
587
00:32:21,814 --> 00:32:24,984
immediately, what are your crown jewels?
588
00:32:24,984 --> 00:32:27,612
And you have an eye on these controls.
589
00:32:27,612 --> 00:32:28,613
Anomaly detection.
590
00:32:28,613 --> 00:32:32,533
We are lucky, kind of lucky today, to
inject a little bit of positivity right.
591
00:32:32,575 --> 00:32:37,830
So within vibe coding I can code within
592
00:32:38,247 --> 00:32:42,543
a very short time a little agent that does
593
00:32:42,543 --> 00:32:46,422
some additional anomaly detection
on some logs that I have.
594
00:32:46,464 --> 00:32:49,384
If I haven't done this before,
I can do this.
595
00:32:49,384 --> 00:32:53,888
One of my favorite topics
is low hanging fruit in security.
596
00:32:53,888 --> 00:32:56,683
So the problem
of a really sophisticated attacker,
597
00:32:56,683 --> 00:32:59,894
and especially with
AI powered and automated attacks.
598
00:33:00,478 --> 00:33:02,730
Their problem is
sometimes that they are too good
599
00:33:02,730 --> 00:33:04,732
and that's where we can catch them.
600
00:33:04,732 --> 00:33:08,111
So things like honeypots and canaries
601
00:33:08,111 --> 00:33:11,114
and anomaly detection which is very basic.
602
00:33:11,155 --> 00:33:15,618
Unlike stolen money, stolen information
doesn't have to be spent immediately.
603
00:33:15,743 --> 00:33:20,123
It can sit quietly for months or years,
604
00:33:20,289 --> 00:33:24,585
waiting until circumstances
change, waiting until a forgotten API
605
00:33:24,585 --> 00:33:29,465
key becomes valuable again, waiting
until someone else discovers
606
00:33:29,465 --> 00:33:33,970
a use for information that seemed
insignificant when it was taken.
607
00:33:33,970 --> 00:33:37,265
Just look for the list and describe
608
00:33:37,432 --> 00:33:41,519
and WHOAMI events in production accounts
that the production
609
00:33:41,519 --> 00:33:45,023
application never uses, of course,
because they are benign from seem.
610
00:33:45,023 --> 00:33:48,026
The seem would say oh,
this is just list user.
611
00:33:48,026 --> 00:33:48,985
Why is this?
612
00:33:48,985 --> 00:33:51,237
List user is not a virus, not a threat.
613
00:33:51,237 --> 00:33:53,322
Oh yes it is!
614
00:33:53,322 -->
00:33:57,744
I believe a security architect today
needs to get back to hands on keyboard
615
00:33:57,744 --> 00:34:02,040
coding, to build this helper application
that can be scrappy too
616
00:34:02,081 --> 00:34:05,334
sometimes,
to pick these low hanging fruits.
617
00:34:05,334 --> 00:34:09,297
Because they're so ridiculously effective
and value for money
618
00:34:09,338 --> 00:34:13,051
and security is not really known
for good ROI typically.
619
00:34:13,092 --> 00:34:16,846
And then in October,
the waiting ended, a group calling itself
620
00:34:16,888 --> 00:34:21,809
Scattered Lapsus$ Hunters launched an
extortion site threatening organizations
621
00:34:21,809 --> 00:34:27,065
reportedly affected by several Salesforce
campaigns, including the Drift breach.
622
00:34:27,231 --> 00:34:30,568
The name Shiny Hunters appeared
among the identities behind it,
623
00:34:30,568 --> 00:34:34,197
but in cybercrime, names are often
just masks.
624
00:34:34,322 --> 00:34:39,202
No public evidence established
whether these were the original UNC6395
625
00:34:39,243 --> 00:34:44,582
attackers, their partners or criminals
who had obtained the stolen data later.
626
00:34:44,749 --> 00:34:49,545
What was clear was that the information
had found a new purpose.
627
00:34:49,587 --> 00:34:54,592
First, it had been searched for keys,
now it was being used as leverage.
628
00:34:54,801 --> 00:34:57,762
The breach had not ended
when the tokens were revoked,
629
00:34:57,762 --> 00:35:00,973
it had simply entered its next phase.
630
00:35:01,140 --> 00:35:05,061
One of the really unsettling
parts of this story
631
00:35:05,061 --> 00:35:08,231
is that the attackers weren't
just stealing data, of course,
632
00:35:08,231 --> 00:35:12,151
they were looking for passwords
and API keys or other digital keys.
633
00:35:12,151 --> 00:35:15,071
Once a trusted integration
has been compromised,
634
00:35:15,071 --> 00:35:20,743
how do you even begin figuring out
where those stolen keys might lead?
635
00:35:20,785 --> 00:35:23,788
That is very difficult
because it's not under your control.
636
00:35:23,830 --> 00:35:25,706
This is actually the scary part.
637
00:35:25,706 --> 00:35:28,584
Lack of visibility
control leads to uncertainty.
638
00:35:28,584 --> 00:35:33,422
Even if I know what happens, I still don't
know the things that I need to know.
639
00:35:33,464 --> 00:35:35,133
What other parts were affected?
640
00:35:35,133 --> 00:35:38,970
You couldn't ask Salesloft to say,
can you give me all data
641
00:35:39,011 --> 00:35:44,851
about any user and integration
usage patterns of the last five days?
642
00:35:44,892 --> 00:35:45,643
Good luck, right?
643
00:35:45,643 --> 00:35:46,686
In the months that followed,
644
00:35:46,686 --> 00:35:50,731
security teams around the world
began asking different questions.
645
00:35:50,773 --> 00:35:53,818
Not simply
have we patched our systems, but
646
00:35:53,860 --> 00:35:57,238
who already has permission to enter them.
647
00:35:57,238 --> 00:35:58,823
So you don't have that data.
648
00:35:58,823 --> 00:36:01,701
That means we have another lack of trust.
649
00:36:01,701 --> 00:36:06,831
So there are other companies
who were victims, who have other customers
650
00:36:06,831 --> 00:36:10,877
who don't know
that you were part of this breach.
651
00:36:11,085 --> 00:36:13,921
Your own intransparenciation leads
652
00:36:13,921 --> 00:36:17,425
to another erosion of trust
because you leave your customers.
653
00:36:17,425 --> 00:36:22,597
I mean, how can 1000-1500 companies
be affected and only 50 come out?
654
00:36:22,680 --> 00:36:25,308
They all have customers
who are in the dark.
655
00:36:25,308 --> 00:36:28,311
But the list of important questions grows,
656
00:36:28,352 --> 00:36:32,648
like which integrations still need
the access they were granted years ago?
657
00:36:32,648 --> 00:36:35,985
Which support platforms
have quietly become repositories
658
00:36:35,985 --> 00:36:39,780
of passwords, of API keys of certificates
and cloud credentials?
659
00:36:39,780 --> 00:36:44,076
And which trusted applications
have become so familiar
660
00:36:44,368 --> 00:36:47,163
that no one remembers
they're even there anymore?
661
00:36:47,163 --> 00:36:48,873
I think this SaaS-to-Saas Proxy
662
00:36:48,873 --> 00:36:51,500
can also contribute
to soften that problem a little bit.
663
00:36:51,500 --> 00:36:54,503
How do you think the CISOs
who will come out on the right side
664
00:36:54,503 --> 00:36:57,965
of history are thinking differently
than, say, perhaps the majority?
665
00:36:58,007 --> 00:37:01,135
There are CISOs and CISOs, so
security directors and security directors.
666
00:37:01,135 --> 00:37:04,222
There are thought leaders
and they are the ones with passion,
667
00:37:04,222 --> 00:37:07,642
and there are the ones
who accidentally ended up there.
668
00:37:07,683 --> 00:37:10,228
They never got into security
because of passion.
669
00:37:10,228 --> 00:37:14,023
And this is where the biggest differences
are, because the ones with passion
670
00:37:14,023 --> 00:37:15,691
- much harder to satisfy.
671
00:37:15,691 --> 00:37:18,736
They asked the next question and
the next question and the next question.
672
00:37:18,736 --> 00:37:20,112
They think of the box.
673
00:37:20,112 --> 00:37:23,157
They are typically also more
interconnected in the community.
674
00:37:23,199 --> 00:37:24,867
The ones who see it as a career path,
675
00:37:24,867 --> 00:37:28,621
and it depends on industry,
it depends on background of the person.
676
00:37:28,913 --> 00:37:31,832
There are also differences
in terms of countries, especially in Asia
677
00:37:31,832 --> 00:37:33,793
where there is more mature,
less mature IT.
678
00:37:33,793 --> 00:37:36,545
Three years ago,
no company in Taiwan had a CISO.
679
00:37:36,545 --> 00:37:38,839
When you have passion, then you learn.
680
00:37:38,839 -->
00:37:40,424
There are many differences
681
00:37:40,424 --> 00:37:43,177
that we can discuss
between different people and organizations
682
00:37:43,177 --> 00:37:45,137
and if they do a better
or not so good job.
683
00:37:45,137 --> 00:37:50,559
But when you work it backwards - why why
why why - the last why for me is passion.
684
00:37:50,601 --> 00:37:53,604
The Drift breach didn't
expose a flaw in Salesforce.
685
00:37:53,604 --> 00:37:57,483
It didn't reveal a catastrophic failure
inside Google Workspace.
686
00:37:57,483 --> 00:38:02,238
Instead, it illuminated something
far more uncomfortable.
687
00:38:02,613 --> 00:38:07,785
Modern enterprises are built less on walls
than they are in relationships.
688
00:38:07,910 --> 00:38:10,204
Applications trust other applications.
689
00:38:10,204 --> 00:38:11,706
Clouds trust services.
690
00:38:11,706 --> 00:38:13,833
Services trust identities.
691
00:38:13,833 --> 00:38:17,837
Break one trusted link
and every connection beyond it deserves
692
00:38:17,837 --> 00:38:19,547
to be questioned.
693
00:38:19,547 --> 00:38:21,882
The attackers didn't steal trust.
694
00:38:21,882 --> 00:38:23,801
They inherited it.
695
00:38:23,801 --> 00:38:29,223
Too often the attackers understand that
before the defenders do.
696
00:38:29,265 --> 00:38:33,311
Today, defenders
rely on attackers making mistakes.
697
00:38:33,311 --> 00:38:36,564
What happens when AI gives an attacker
the ability
698
00:38:36,564 --> 00:38:40,776
to learn from those mistakes
and change its behavior in real time?
699
00:38:40,776 --> 00:38:41,986
It's already reality.
700
00:38:41,986 --> 00:38:44,739
There's four types of AI: AI in security
- we're getting attacked,
701
00:38:44,739 --> 00:38:47,783
we need to defend; AI
in security compliant adoption;
702
00:38:47,825 --> 00:38:51,120
AI in the software development
phase - we're using AI to build AI,
703
00:38:51,120 --> 00:38:54,165
and then we run AI stuff;
And then AI in the supply chain.
704
00:38:54,165 --> 00:38:56,000
So we're talking about the first one.
705
00:38:56,000 --> 00:38:57,418
We're getting attacked by AI.
706
00:38:57,418 --> 00:38:59,003
The problem stays - the controls.
707
00:38:59,003 --> 00:39:01,047
And what we need to do,
what we need to think about in
708
00:39:01,047 --> 00:39:02,673
these four categories are different.
709
00:39:02,673 --> 00:39:07,053
What actually happened
since AI is supercharging threat actors?
710
00:39:07,053 --> 00:39:10,639
We need to assume that our controls,
our protective controls
711
00:39:10,639 --> 00:39:14,101
on the network identity and application
level are less effective.
712
00:39:14,143 --> 00:39:16,604
Let's stop freaking out,
keep calm and carry on.
713
00:39:16,604 --> 00:39:17,938
It's just a new normal.
714
00:39:17,938 --> 00:39:21,901
We pivot to an assumed state of compromise
and need to use detection
715
00:39:21,901 --> 00:39:24,945
and remediation
to play a more important
role.
716
00:39:24,945 --> 00:39:28,366
If one protection goes down,
we need to ramp up the other side.
717
00:39:28,366 --> 00:39:33,454
The new skill is to understand
where is AI not perfect?
718
00:39:33,454 --> 00:39:36,791
Context sensitive anomaly detection
gives us a very high chance,
719
00:39:36,791 --> 00:39:39,335
no matter how sophisticated
the threat actor is.
720
00:39:39,335 --> 00:39:43,172
Zero Trust network access,
which is part of the Zero Trust paper,
721
00:39:43,214 --> 00:39:45,216
means you need to establish trust first.
722
00:39:45,216 --> 00:39:47,760
Otherwise you don't even see that network.
723
00:39:47,760 --> 00:39:48,469
Hang on a second.
724
00:39:48,469 --> 00:39:50,262
If I can't see anything, I can't scan it.
725
00:39:50,262 --> 00:39:51,806
If I can't scan it, I can't feed my AI.
726
00:39:51,806 --> 00:39:53,474
So AI is becoming useless.
727
00:39:53,474 --> 00:39:55,226
I need to solve this problem first.
728
00:39:55,226 --> 00:39:58,521
I need to get to something
that I can actually attack.
729
00:39:58,521 --> 00:40:03,984
And when I use a non-fishable
second factor, then with even
730
00:40:03,984 --> 00:40:07,071
all the power of AI, you raise the bar
731
00:40:07,071 --> 00:40:10,199
so high that another game kicks
732
00:40:10,199 --> 00:40:14,578
in that we are always playing:
Who runs the fastest away from the bear.
733
00:40:14,578 --> 00:40:19,291
If I run faster, that means
that it is harder for the attacker.
734
00:40:19,291 --> 00:40:20,459
If it's harder for the attacker -
735
00:40:20,459 --> 00:40:25,047
they run a business too
and we can use AI to run much faster.
736
00:40:25,047 --> 00:40:26,757
Let's not be scared about AI.
737
00:40:26,757 --> 00:40:32,012
Let's embrace AI to build agile defenses
that operate at machine speed,
738
00:40:32,012 --> 00:40:35,141
because if we like it or not,
that's the game we are in.
739
00:40:35,182 --> 00:40:39,895
The Five Eyes, recently issued
an unusually urgent
740
00:40:39,937 --> 00:40:41,981
warning about AI and cybersecurity.
741
00:40:41,981 --> 00:40:45,192
And when you combine that warning
with attacks like Drift
742
00:40:45,192 --> 00:40:48,988
and the conversations
you're having with security leaders today,
743
00:40:48,988 --> 00:40:52,241
what do you think the biggest assumption
about enterprise security
744
00:40:52,450 --> 00:40:56,454
you think CISOs need to leave behind,
let go of?
745
00:40:56,454 --> 00:41:00,374
I believe it's
now time to let go of our old security
746
00:41:00,416 --> 00:41:03,252
strategies,
because most security strategies,
747
00:41:03,252 --> 00:41:06,422
and therefore security enablement programs
748
00:41:06,505 --> 00:41:09,592
were written pre-AI, pre-Mythos.
749
00:41:09,633 --> 00:41:10,968
These are the three things:
750
00:41:10,968 --> 00:41:15,598
Think out of the box; Aim for resilience
not compliance; And no marketecture.
751
00:41:15,681 --> 00:41:17,892
So
spend your money wisely.
752
00:41:17,892 --> 00:41:21,479
That little chat window still
sits in the corner of countless websites.
753
00:41:21,479 --> 00:41:26,025
It still asks the same polite question:
How can we help you today?
754
00:41:26,150 --> 00:41:28,819
For most visitors, it's
the beginning of a conversation.
755
00:41:28,819 --> 00:41:32,448
And for security leaders,
it should also be a reminder.
756
00:41:32,490 --> 00:41:36,785
Sometimes the most dangerous door
in an organization is the one
757
00:41:36,785 --> 00:41:39,580
no one remembers installing.
Thank you so much!
758
00:41:39,580 --> 00:41:40,831
It was great having you on the show.
759
00:41:40,831 --> 00:41:44,335
It was a fantastic conversation
and I hope to have you back again soon.
760
00:41:44,376 --> 00:41:46,587
It was a great pleasure.
Thank you very much, Jeremy.
761
00:41:46,587 --> 00:41:47,463
That was really good.
762
00:41:47,463 --> 00:41:50,049
I really enjoyed it.
Thank you very much for the opportunity.
763
00:41:51,300 --> 00:41:54,470
Centuries ago, the most carefully defended
castles
764
00:41:54,512 --> 00:41:58,516
were surrounded by thick walls,
tall towers and guarded gates,
765
00:41:58,516 --> 00:42:02,728
but many also contained
something called a postern.
766
00:42:02,978 --> 00:42:07,316
A smaller entrance built into the wall,
so servants and messengers and supplies
767
00:42:07,316 --> 00:42:11,070
could easily pass without the trouble
of opening the main gate.
768
00:42:11,362 --> 00:42:14,114
But every door in a castle wall represents
769
00:42:14,114 --> 00:42:17,076
a potential weakness, an attack surface.
770
00:42:17,076 --> 00:42:20,079
Today,
the castle has become an enterprise.
771
00:42:20,079 --> 00:42:23,332
The postern is a support portal
or a chatbox,
772
00:42:23,332 --> 00:42:27,378
or an integration approved years ago
and quietly forgotten.
773
00:42:27,419 --> 00:42:29,713
On its own,
it isn't malicious or sinister.
774
00:42:29,713 --> 00:42:33,342
It simply performs its assigned task,
allowing information
775
00:42:33,342 --> 00:42:38,597
to conveniently travel through the wall
until someone else discovers
776
00:42:38,597 --> 00:42:42,810
that convenience, that trust,
and finds a way to weaponize it.
777
00:42:43,143 --> 00:42:46,146
The danger was never that organizations
chose to trust.
778
00:42:46,146 --> 00:42:47,898
Without trust, modern business stops.
779
00:42:47,898 --> 00:42:51,777
The danger begins
when trust outlives scrutiny,
780
00:42:51,819 --> 00:42:55,197
when yesterday's permission becomes
today's open door.
781
00:42:55,364 --> 00:42:58,701
Because somewhere at this very moment,
a forgotten
782
00:42:58,701 --> 00:43:01,912
connection
is still working exactly as designed.
783
00:43:02,204 --> 00:43:05,207
The only question is for whom?
784
00:43:05,416 --> 00:43:08,752
And so we must remain forever vigilant
785
00:43:08,752 --> 00:43:13,048
and always listening for The CISO Signal.
786
00:43:14,049 --> 00:43:14,508
If you
787
00:43:14,508 --> 00:43:17,678
enjoyed this episode, please like,
share and subscribe.
788
00:43:17,678 --> 00:43:19,930
If you didn't, thanks for listening
this long.
789
00:43:19,930 --> 00:43:23,475
We'll see you
on the next episode of The CISO Signal.
790
00:43:23,559 --> 00:43:24,810
All episodes are based
791
00:43:24,810 --> 00:43:29,023
on publicly available
reports, postmortems, and expert analysis.
792
00:43:29,106 --> 00:43:32,568
While we've done our best
to ensure accuracy, some cybersecurity
793
00:43:32,568 --> 00:43:36,655
incidents evolve over time
and not all details have been confirmed.
794
00:43:36,905 --> 00:43:42,119
Our goal is to inform and entertain,
not to assign blame.
795
00:43:42,328 --> 00:43:43,746
Where facts are unclear,
796
00:43:43,746 --> 00:43:48,000
we've used cautionary language
and we always welcome your corrections.
797
00:43:48,042 --> 00:43:50,586
Thanks for listening to The CISO Signal.