WEBVTT

00:00:00.100 --> 00:00:03.780
A disturbing,
yet not really surprising article or

00:00:03.820 --> 00:00:07.280
post was published by Anthropic yesterday.

00:00:07.360 --> 00:00:11.080
An article about an AI-orchestrated

00:00:11.240 --> 00:00:14.760
cyber espionage campaign, a cyberattack

00:00:15.060 --> 00:00:17.720
carried out with help of AI,
with help of Claude code.

00:00:17.740 --> 00:00:21.240
And it's an interesting article.
I already read it

00:00:21.280 --> 00:00:25.120
together with you here. So Anthropic,
in this article,

00:00:25.160 --> 00:00:28.840
describes a, a cyberattack on

00:00:28.900 --> 00:00:32.740
various companies that
was carried out almost entirely with

00:00:32.759 --> 00:00:36.740
help of AI,
almost entirely with help of Claude code,

00:00:36.780 --> 00:00:40.580
by jailbreaking Claude code,
by getting it to do

00:00:40.880 --> 00:00:44.610
things it normally shouldn't do.
And let's take a closer look at

00:00:44.680 --> 00:00:48.600
how that played out.
"In mid-September 2025,

00:00:48.620 --> 00:00:52.510
we, Anthropic,
detected suspicious activity that later,

00:00:52.540 --> 00:00:56.500
investigation determined to be a highly
sophisticated espionage campaign.

00:00:56.540 --> 00:00:59.980
The attackers used AI's agentic
capabilities to an

00:01:00.060 --> 00:01:04.040
unprecedented degree,
using AI not just as an advisor, but to

00:01:04.120 --> 00:01:07.830
execute the cyberattacks themselves."
And that's, that

00:01:07.900 --> 00:01:11.260
is huge that you can

00:01:11.360 --> 00:01:14.900
already use AI, AI that is

00:01:14.920 --> 00:01:18.720
published by Anthropic,
AI models by Anthropic, and the

00:01:18.780 --> 00:01:22.540
Claude code tool by Anthropic to carry out

00:01:22.600 --> 00:01:25.040
cyberattacks. And, uh, here's why that is

00:01:25.060 --> 00:01:29.040
important.
Today we're living in a world where

00:01:29.100 --> 00:01:32.540
the most capable models are the

00:01:32.600 --> 00:01:35.880
models by OpenAI, Anthropic, X,

00:01:35.940 --> 00:01:39.860
Google,
and of course we've got some very capable

00:01:39.920 --> 00:01:43.690
too,
but we're still at a point in time on a,

00:01:43.760 --> 00:01:47.380
timeline, we are still at a point in time

00:01:47.420 --> 00:01:50.960
where most of these models and the,
the most capable

00:01:51.020 --> 00:01:54.620
models are controlled by companies,

00:01:55.120 --> 00:01:57.840
most of them by companies in democracies.

00:01:57.900 --> 00:02:01.880
Now, I'll not say that these companies
are saints or

00:02:02.280 --> 00:02:06.120
don't do bad stuff, but they're not bad

00:02:06.240 --> 00:02:09.560
actors in the sense of this article.

00:02:09.580 --> 00:02:12.590
They're not cyber, um, attackers

00:02:12.840 --> 00:02:16.460
obviously. However, in the future,
in the not too

00:02:16.600 --> 00:02:19.840
distant future, we will be at a place

00:02:19.900 --> 00:02:23.410
where these models,
these very capable models

00:02:24.020 --> 00:02:27.750
will also be owned by bad actors

00:02:27.780 --> 00:02:31.600
themselves.
So right now here in this article,

00:02:31.680 --> 00:02:35.200
about a cyberattack that
was carried out with help of

00:02:35.280 --> 00:02:39.000
Anthropic and Claude code.
And it's bad enough that this is possible,

00:02:39.160 --> 00:02:43.060
trick those tools
and models into doing stuff they shouldn't

00:02:43.100 --> 00:02:46.720
we'll get back to how that happened,
of course, but that is

00:02:46.760 --> 00:02:50.560
today. We still have to apply tricks,

00:02:50.640 --> 00:02:53.860
uh, to abuse these models. In a

00:02:53.940 --> 00:02:57.480
future,
these models will simply belong to the bad

00:02:57.660 --> 00:03:01.420
actors themselves.
There will be open models that are capable

00:03:01.480 --> 00:03:05.350
enough of doing that,
so then certain control mechanisms,

00:03:05.400 --> 00:03:09.140
get back here in this article,
won't even be there anymore.

00:03:09.200 --> 00:03:13.000
We'll be in a future where bad actors have

00:03:13.040 --> 00:03:16.970
direct,
uncontrolled access to very capable models

00:03:17.140 --> 00:03:20.940
can be fine-tuned for their purposes,
that can be

00:03:21.040 --> 00:03:24.900
trained for their purposes,
that can use tools that were

00:03:25.000 --> 00:03:27.500
purpose built for malicious stuff.

00:03:27.580 --> 00:03:31.380
That is where we're heading to,
and even today where we're not there

00:03:31.520 --> 00:03:35.040
yet or where this is still a niche,
even today, we

00:03:35.200 --> 00:03:39.190
have fully or almost fully AI-controlled,

00:03:39.420 --> 00:03:43.300
uh, cyberattacks. So this
is definitely a scary article

00:03:43.360 --> 00:03:47.020
and, and a scary future,
which makes it very clear that

00:03:47.120 --> 00:03:49.560
cybersecurity and

00:03:49.620 --> 00:03:53.400
preventing attacks will be a super

00:03:53.520 --> 00:03:57.420
big challenge. It always has been,
but now with AI where everything can be

00:03:57.520 --> 00:04:00.840
quicker and more automated
and harder to trace back to

00:04:00.880 --> 00:04:04.640
individuals,
it will be an even more important

00:04:04.660 --> 00:04:07.840
topic. But back to this article.
"The threat

00:04:07.940 --> 00:04:11.760
actor,
whom we assess with high confidence

00:04:11.980 --> 00:04:15.560
state-sponsored group,
manipulated our Claude code

00:04:15.980 --> 00:04:19.600
tool into attempting infiltration into
roughly 30

00:04:19.640 --> 00:04:23.630
global targets and succeeded in a small
number of cases." So it was not just

00:04:23.660 --> 00:04:27.500
an attempt. They succeeded. "The operation

00:04:27.540 --> 00:04:31.120
targeted large tech companies,
financial institutions,

00:04:31.160 --> 00:04:33.810
companies, and government agencies.

00:04:34.020 --> 00:04:37.580
Uh, we believe this
is the first documented case of a

00:04:37.640 --> 00:04:41.420
cyberattack executed without substantial
human intervention." This is so

00:04:41.520 --> 00:04:43.860
big, without substantial human

00:04:43.920 --> 00:04:47.690
intervention. And again, that
is why this is

00:04:47.740 --> 00:04:51.540
such a scary future where bad actors don't
even

00:04:51.600 --> 00:04:55.260
need to rely on, um, these controlled

00:04:55.520 --> 00:04:58.740
AI models, uh,
which they still have to rely on today.

00:04:58.750 --> 00:05:01.969
"Upon detecting this activity,
we immediately launched an investigation

00:05:02.020 --> 00:05:06.000
understand its scope and nature. Uh,
over the following 10 days, as we

00:05:06.020 --> 00:05:09.960
mapped the severity
and full extent of the operation,

00:05:09.980 --> 00:05:13.840
they were identified." So again,
these were really regular

00:05:13.940 --> 00:05:16.840
Claude accounts. These people were using

00:05:16.900 --> 00:05:19.910
Claude, the model hosted by Anthropic.

00:05:19.920 --> 00:05:23.640
They did not kind of steal it,
run it on their own servers

00:05:23.740 --> 00:05:27.560
models.
They used the models you can use too via

00:05:27.600 --> 00:05:31.420
via Claude code.
This campaign has substantial

00:05:31.440 --> 00:05:35.280
implications for cybersecurity in the age
of AI agents, as I just said,

00:05:35.340 --> 00:05:39.290
because everything can be automated
and it's already possible today, uh,

00:05:39.340 --> 00:05:42.090
where there are guardrails in place.

00:05:42.100 --> 00:05:45.800
And again, think of
that future where we have no guardrails

00:05:45.920 --> 00:05:49.880
actors don't have guardrails. "Uh,
these attacks are likely

00:05:49.920 --> 00:05:51.900
to only grow in their effectiveness.

00:05:51.940 --> 00:05:55.680
To keep pace with this rapidly advancing
threat, we have expanded our

00:05:55.700 --> 00:05:59.650
detection capabilities
and developed better classifiers to flag

00:05:59.720 --> 00:06:03.472
activities." And that's the important part
here.This is what

00:06:03.482 --> 00:06:06.832
Anthropic is trying to do today to make
sure that their

00:06:06.872 --> 00:06:10.742
models can't be abused for malicious
tasks,

00:06:10.812 --> 00:06:14.612
that Claude code can't be abused
and ultimately,

00:06:14.632 --> 00:06:17.952
their APIs,
which Claude code uses in the end.

00:06:18.012 --> 00:06:21.972
This all won't matter at all in the future
because

00:06:22.052 --> 00:06:25.912
in a future where bad actors themselves
have their own models

00:06:25.972 --> 00:06:29.762
running on their own servers,
these guardrails won't

00:06:29.792 --> 00:06:31.902
matter. Well,
obviously they will still matter.

00:06:31.972 --> 00:06:35.552
Obviously,
you still want to make it easier than it's

00:06:35.632 --> 00:06:39.072
obviously not all bad actors will have
their

00:06:39.152 --> 00:06:42.912
own malicious models, but especially
if we're talking about

00:06:42.932 --> 00:06:46.512
state-controlled bad actors or big

00:06:46.572 --> 00:06:50.352
cyber attacker groups. Let's, let's be

00:06:50.372 --> 00:06:54.252
real.
Of course they will have access to their

00:06:54.292 --> 00:06:57.882
on their own servers,
so this will not matter at

00:06:57.992 --> 00:07:01.932
all in that future. Obviously,
it will matter still because you don't

00:07:01.972 --> 00:07:05.832
make it easier and it will at least filter
out a significant

00:07:05.932 --> 00:07:09.812
group of potential bad actors
that don't have access to their own

00:07:09.872 --> 00:07:13.852
models. So yeah, it's important,
but it will not be enough.

00:07:13.912 --> 00:07:17.712
Companies themselves need to ramp up

00:07:17.772 --> 00:07:21.732
their cybersecurity game, which
is easier said than done because

00:07:21.772 --> 00:07:25.752
it has been true for the last 10 years,
of course, even without AI, but it's

00:07:25.772 --> 00:07:28.872
becoming even more important in the age of
AI.

00:07:30.132 --> 00:07:33.612
So yeah, that is, that
is the big problem here.

00:07:33.652 --> 00:07:37.032
Now, let's see how the, uh,
cyberattack worked.

00:07:37.132 --> 00:07:40.762
Uh,
the attack relied on several features of

00:07:40.792 --> 00:07:44.432
exist or were in much more nascent form
just a year ago.

00:07:44.452 --> 00:07:48.172
Intelligence:
Models' general levels of capability have

00:07:48.272 --> 00:07:52.152
point that they can follow complicance-
complex instructions and understand

00:07:52.192 --> 00:07:56.012
context in ways that make very
sophisticated tasks possible.

00:07:56.052 --> 00:07:59.722
Not only that,
but several of their well-developed

00:07:59.752 --> 00:08:03.212
particular software coding,
lend themselves to being used in

00:08:03.252 --> 00:08:07.122
Sure, because now with models that
are smarter,

00:08:07.152 --> 00:08:10.782
and just to be very clear here,
we're still talking about models that

00:08:10.892 --> 00:08:14.872
just generate tokens,
but of course by generating these tokens

00:08:14.932 --> 00:08:18.922
that are able, they
are able to describe the usage of tools

00:08:18.952 --> 00:08:22.452
tools, they become more capable.
With all the

00:08:22.472 --> 00:08:26.092
fine-tuning they received,
they also generate more tokens that are

00:08:26.352 --> 00:08:30.332
likely to be or more likely to be the
tokens you want to generate, so that

00:08:30.372 --> 00:08:34.292
is what intelligence means here.
They're not really intelligent, but they

00:08:34.352 --> 00:08:38.182
have been tuned especially for software
development such that

00:08:38.212 --> 00:08:41.962
they are much more likely to generate
meaningful output

00:08:42.272 --> 00:08:46.092
and especially also output
that allows them to describe tool

00:08:46.172 --> 00:08:49.892
use and then use those tools,
so execute code

00:08:49.902 --> 00:08:53.732
that does something. Uh, for example,
send an HTTP request and so

00:08:53.772 --> 00:08:57.572
on, and it's that combination
that makes them more capable

00:08:57.912 --> 00:09:01.112
in the end. And of course, yeah,
that's exactly what you need for automated

00:09:01.132 --> 00:09:05.012
cyberattacks,
because you need a model that's able to

00:09:05.072 --> 00:09:07.272
follow your instructions related to that.

00:09:07.312 --> 00:09:10.912
You need a model that's able to send HTTP
requests, phishing

00:09:10.972 --> 00:09:14.662
emails, whatever, and all that
is what these models can do

00:09:14.752 --> 00:09:18.242
quite well in the end. Agency:
Models can act as agents.

00:09:18.272 --> 00:09:21.872
That is,
they can run in loops where they take

00:09:21.932 --> 00:09:24.812
tasks and make decisions with only minimal
occasional human

00:09:24.872 --> 00:09:27.432
input. That's another important step.

00:09:27.512 --> 00:09:31.242
This is what allows these, um, systems
or these

00:09:31.272 --> 00:09:35.192
attacks here in this case to work with
only minimal human input

00:09:36.132 --> 00:09:40.032
because these models and the software
that uses

00:09:40.072 --> 00:09:43.712
these models can go for so much longer
and it's important to

00:09:43.752 --> 00:09:47.672
differentiate here. The model is,
is still just the

00:09:47.712 --> 00:09:51.332
thing that receives a prompt
and sends back some tokens.

00:09:51.352 --> 00:09:55.292
That, that has not changed.
But it's the software, Claude Code, for

00:09:55.372 --> 00:09:59.252
example, that then takes that output
and sends back another

00:09:59.312 --> 00:10:02.742
message to the same API with that output,
with the original

00:10:02.852 --> 00:10:06.572
task, with some meta instructions like,
"Please check if that

00:10:06.612 --> 00:10:09.412
output answers the question by the user.

00:10:09.472 --> 00:10:12.932
You got these tools available,
please tell me if you want to use a tool."

00:10:12.992 --> 00:10:16.912
how the software around the models in the
end makes these

00:10:16.992 --> 00:10:20.711
models more capable,
not because the model does everything on

00:10:20.752 --> 00:10:24.352
the model is capable of giving the
software the

00:10:24.392 --> 00:10:28.372
result it needs.
The software then feeds these enriched

00:10:28.412 --> 00:10:32.112
the model and it's this loop
that keeps the whole system going and that

00:10:32.192 --> 00:10:35.972
leads to these agentic systems that can go

00:10:36.052 --> 00:10:39.752
on for longer, that can use tools and
that require less

00:10:39.812 --> 00:10:43.692
human input. And yeah, tools, that
is therefore the other missing piece

00:10:43.732 --> 00:10:46.972
here, of course,
that models have access to a wide array of

00:10:47.012 --> 00:10:50.612
tools. They can now search the web,
retrieve data, perform many other

00:10:50.692 --> 00:10:53.662
actions that were previously the sole
domain of human operators.

00:10:53.692 --> 00:10:57.492
In the case of cyberattacks,
the tools might include password crackers,

00:10:57.552 --> 00:11:00.392
scanners and other security-related
software.

00:11:00.412 --> 00:11:04.152
Because again,
it's not all just GitHub MCPs.

00:11:04.192 --> 00:11:07.692
It can be all kind of tools you could, uh,
expose

00:11:07.752 --> 00:11:11.572
to your, um, model or to the software
that uses

00:11:11.612 --> 00:11:14.332
these models and
that runs these agentic tasks.

00:11:14.412 --> 00:11:18.152
So they got a nice diagram in this
article, but in the end

00:11:18.252 --> 00:11:21.721
the attack played out relatively, uh,

00:11:21.771 --> 00:11:25.122
simple. They,
they describe it in greater detail down

00:11:25.152 --> 00:11:29.012
there. They convinced Claude Code to do

00:11:29.031 --> 00:11:31.192
stuff it normally shouldn't be able to do.

00:11:31.232 --> 00:11:34.781
They had to convince Claude, which
is extensively trained to avoid harmful

00:11:34.852 --> 00:11:36.652
behaviors, to engage in the attack.

00:11:36.672 --> 00:11:40.652
They did so by jailbreaking it,
effectively tricking it to bypass

00:11:40.672 --> 00:11:44.452
its guardrails,
and that's the part where Anthropic

00:11:44.472 --> 00:11:48.332
back better, not just in Claude Code
but in the

00:11:48.452 --> 00:11:52.292
models themselves,
so on their API where they scan all the

00:11:52.332 --> 00:11:56.142
requests that reach their models,
so to say, and take

00:11:56.152 --> 00:11:59.942
better... eh,
they try to do a better job at detecting

00:12:00.031 --> 00:12:04.021
injections in the end because these
attackers broke down their attacks into

00:12:04.112 --> 00:12:07.962
small, seemingly innocent tasks
that Claude would execute without being

00:12:07.992 --> 00:12:10.762
provided the full context of their
malicious purpose.

00:12:10.762 --> 00:12:14.712
They also told Claude that it
was an employee of a legitimate

00:12:14.732 --> 00:12:18.632
cybersecurity firm and
was being used in defensive testing.That's

00:12:18.652 --> 00:12:22.432
good old trick. I think that
is how jailbreaking was already done two

00:12:22.492 --> 00:12:25.722
years ago with the early ChatGPT models.

00:12:25.812 --> 00:12:29.692
Eh,
you tell it that you need this information

00:12:29.792 --> 00:12:33.052
and it'll happily expose its system
prompt.

00:12:33.092 --> 00:12:36.952
Kind of a simplification
but that's still how prompt injections can

00:12:36.972 --> 00:12:40.802
days.
That you try to apply various techniques

00:12:40.832 --> 00:12:44.692
are very interesting techniques
when it comes to that,

00:12:44.972 --> 00:12:48.232
eh,
including the use of special tokens you

00:12:48.242 --> 00:12:51.372
message to tr- to get the

00:12:51.472 --> 00:12:54.822
model to generate output it normally
shouldn't

00:12:54.872 --> 00:12:57.972
generate.
The attackers then initiated the second

00:12:58.092 --> 00:13:01.241
involved Claude code ins-
inspecting the target organization's

00:13:01.272 --> 00:13:04.692
So yeah,
that's then essentially what Claude code

00:13:04.732 --> 00:13:08.652
did. Then with minimal human input, um, it

00:13:08.732 --> 00:13:11.972
is in the end then used its agentic
capabilities, its

00:13:12.032 --> 00:13:15.741
tools, to really, um,

00:13:15.832 --> 00:13:19.812
scan networks, write code, and do all that

00:13:19.822 --> 00:13:23.702
stuff without a human telling it exactly
what to do.

00:13:23.752 --> 00:13:27.692
So, it was, as mentioned earlier, uh,
a fully or almost fully

00:13:27.752 --> 00:13:31.292
automated attack. So,
in the next phases of the attack, Claude

00:13:31.372 --> 00:13:35.262
identified and tested security
vulnerabilities in the target

00:13:35.352 --> 00:13:38.392
systems by researching
and writing its own exploit

00:13:38.512 --> 00:13:42.252
code. Having done so, the framework
was able to use Claude to

00:13:42.292 --> 00:13:45.952
harvest credentials, usernames
and passwords that allowed it to further

00:13:45.972 --> 00:13:48.192
then extract a large amount of private
data.

00:13:48.232 --> 00:13:52.192
So, it did really research,
write the code to

00:13:52.252 --> 00:13:55.982
get into systems, uh,
of other companies and then in those

00:13:56.092 --> 00:13:59.752
systems write more code to extract data,

00:13:59.852 --> 00:14:03.362
um, and- and-
and compromise these systems and- and- and

00:14:03.692 --> 00:14:07.062
do bad stuff in there once it
was in there.

00:14:07.092 --> 00:14:10.952
The highest privileged accounts
were identified, backdoors

00:14:11.012 --> 00:14:14.752
all the stuff that happened after it
was in the systems and data were exfil-

00:14:14.992 --> 00:14:18.292
exfiltrated with minimal human
supervision.

00:14:18.312 --> 00:14:21.692
In a final phase,
the attackers had Claude produce

00:14:21.732 --> 00:14:24.992
the attack,
creating helpful files with the stolen

00:14:25.112 --> 00:14:28.712
analyzed which would assist the framework
in planning the next stage of the

00:14:28.772 --> 00:14:32.732
threats actor, eh,
of the threat actor's cyber operations.

00:14:32.792 --> 00:14:36.412
Overall, the threat actor
was able to use AI to perform 80 to 90% of

00:14:36.432 --> 00:14:40.292
campaign with human intervention required
only sporadically, perhaps four to

00:14:40.312 --> 00:14:42.942
six critical decision points per hacking
campaign.

00:14:42.992 --> 00:14:46.952
That is nothing. That is nothing. That
is such

00:14:47.032 --> 00:14:50.652
a scale at which you can run these attacks
and again, especially in a

00:14:50.732 --> 00:14:54.552
future where you don't have to work
against certain guardrails, where you can

00:14:54.712 --> 00:14:58.552
just focus on getting the job done
and you have to spend

00:14:58.612 --> 00:15:00.462
energy on getting around guardrails.

00:15:00.512 --> 00:15:03.912
That is really a scary future.
This degree of

00:15:03.992 --> 00:15:07.892
automation is really, really, uh, scary

00:15:07.902 --> 00:15:09.902
here. Claude didn't always work perfectly.

00:15:09.912 --> 00:15:13.372
It occasionally hallucinated credentials
or claimed to have extracted secret

00:15:13.392 --> 00:15:15.742
information that was in fact publicly
available.

00:15:15.792 --> 00:15:19.782
This remains an obstacle to fully, uh,
autonomous cyber attacks and this is

00:15:19.832 --> 00:15:22.912
not just an obstacle for cyber attacks,
this is of course an obstacle for

00:15:22.972 --> 00:15:26.752
everybody,
for us developers too because

00:15:26.812 --> 00:15:30.672
problem and will stay a problem because as
I mentioned before, it's so

00:15:30.732 --> 00:15:33.832
easy to forget but these
are token generation

00:15:34.032 --> 00:15:36.981
machines. Always have been,
always will be.

00:15:36.992 --> 00:15:38.832
The large language models, I mean.

00:15:38.872 --> 00:15:42.772
They are generating tokens and they
are generating

00:15:42.832 --> 00:15:46.772
the most likely token as the next token
based on all the

00:15:46.782 --> 00:15:50.172
tokens that came before it, and that
is something that can and

00:15:50.272 --> 00:15:53.232
always will have the danger of

00:15:53.252 --> 00:15:56.832
hallucinating. So, that of course
is a problem in

00:15:56.872 --> 00:16:00.752
general. Good to see
that it can then also be helpful

00:16:00.832 --> 00:16:04.532
when it comes to defending against
malicious tasks because

00:16:04.572 --> 00:16:08.322
those also are hurt by hallucination
but of course that is a general

00:16:08.322 --> 00:16:11.882
problem, uh, we face, uh,
and it will not be a

00:16:11.912 --> 00:16:14.602
significant, uh, defense mechanism

00:16:14.692 --> 00:16:18.652
unfortunately. Because in the end
if everything's automated, it's

00:16:18.732 --> 00:16:22.322
just a question of scale and
if some attacks fail because of

00:16:22.352 --> 00:16:26.242
hallucination, well, does
that really matter if you can run

00:16:26.252 --> 00:16:28.462
thousands of attacks in parallel?

00:16:28.532 --> 00:16:31.682
I'm not sure it does.
Cybersecurity implications.

00:16:31.712 --> 00:16:35.492
The barriers to performing sophisticated
cyber attacks have dropped

00:16:35.512 --> 00:16:38.731
substantially and we predict
that they'll continue to do so.

00:16:38.752 --> 00:16:42.492
With the correct setup,
threat actors can now use agentic AI

00:16:42.512 --> 00:16:46.312
extended periods to do the work of entire
teams of experienced hackers.

00:16:46.322 --> 00:16:50.152
So kind of the same thing
that applies to normal software

00:16:50.212 --> 00:16:53.892
also be more productive with AI,
and I got a video coming up on

00:16:53.932 --> 00:16:57.632
way, just that that
is the case for malicious tasks,

00:16:57.692 --> 00:17:01.412
even worse because there you don't even
have to

00:17:01.512 --> 00:17:04.642
care about things like code quality.

00:17:04.672 --> 00:17:08.082
Obviously you want to have a successful
attack but in the end if everything's

00:17:08.172 --> 00:17:11.732
automated,
you also can care a lot about just the

00:17:11.832 --> 00:17:15.672
scale.
And if you can automate thousands of

00:17:15.732 --> 00:17:19.472
attacks to run in parallel,
it doesn't really

00:17:19.532 --> 00:17:23.512
matter to you if you might have code
quality problems

00:17:23.552 --> 00:17:27.132
or anything like that. So, uh,

00:17:27.212 --> 00:17:31.132
already with the systems today where you
could argue about potential

00:17:31.252 --> 00:17:35.182
problems they have
when it comes to generating code,

00:17:35.252 --> 00:17:38.682
matter for attacks like this because you
need a result that's just good

00:17:38.732 --> 00:17:42.672
enough. And again, chances
are definitely high that results

00:17:42.712 --> 00:17:46.652
will also get better in the future
and we'll be dealing with systems

00:17:46.662 --> 00:17:50.232
that don't even have guardrails.
And as they say here, less

00:17:50.272 --> 00:17:53.712
experienced and resourced groups can now
potentially perform

00:17:53.732 --> 00:17:56.932
large-scale, uh, attacks of this nature.

00:17:57.002 --> 00:18:00.652
This attack is an escalation even on the
wipe hacking findings we

00:18:00.692 --> 00:18:04.621
reported this summer. In those operations,
humans were much, uh, still in

00:18:04.652 --> 00:18:07.572
the loop, uh, directing the operations.

00:18:07.652 --> 00:18:09.492
Here, human involvement was much less

00:18:09.552 --> 00:18:13.032
frequent. And although we have

00:18:13.052 --> 00:18:16.972
visibility into Claude usage,
this case study probably reflects

00:18:17.012 --> 00:18:20.892
of behavior across frontier AI models
and demonstrates how threat actors

00:18:20.932 --> 00:18:24.702
are adapting their operations. By the way,
this is one case that was

00:18:24.752 --> 00:18:28.625
caught by Anthropic.... not sure
if all the cases are being caught, also

00:18:28.716 --> 00:18:32.656
by Google, um, OpenAI and so on.
This raises

00:18:32.676 --> 00:18:36.176
an important question.
If AI models can be misused for

00:18:36.186 --> 00:18:39.146
scale, why continue to develop
and release them?

00:18:39.156 --> 00:18:43.136
The answer is that very, that very, a-
abilities that allow Claude to be used in

00:18:43.176 --> 00:18:46.005
these attacks also make it crucial for
cyber defense.

00:18:46.035 --> 00:18:49.466
Well, (smacks lips) uh,
that's kind of a weak argument, I'll say,

00:18:49.495 --> 00:18:53.406
because if you have one thing
that makes a problem much

00:18:53.436 --> 00:18:57.326
bigger, saying, "Yeah,
but it can also help with the solution,"

00:18:57.376 --> 00:19:01.116
kind of bad, right? So, uh, I-
I'm not really

00:19:01.176 --> 00:19:04.505
sure about that.
If we would not have these models...

00:19:04.535 --> 00:19:07.636
And just to be clear, that
is not (laughs) something that's going to

00:19:07.656 --> 00:19:11.636
But if we would not have them,
it would probably be better than

00:19:11.716 --> 00:19:15.636
in the context of cyberattacks
and defense because

00:19:15.645 --> 00:19:18.396
will always be one step, uh, behind.

00:19:18.476 --> 00:19:22.416
So, uh,
I definitely see these tools more as an

00:19:22.456 --> 00:19:26.055
advantage for the attackers
and a big disadvantage, uh, for

00:19:26.136 --> 00:19:29.436
the, uh, companies
that have to defend against these attacks.

00:19:29.456 --> 00:19:33.356
So that's kind of a weak argument,
my argument would be, it doesn't

00:19:33.476 --> 00:19:37.015
matter if Anthropic, OpenAI and so on

00:19:37.055 --> 00:19:40.696
continue developing AI models,
and obviously they will, just to be very

00:19:40.755 --> 00:19:44.596
clear.
And there are way more arguments to be

00:19:44.666 --> 00:19:48.275
cybersecurity. This is just one very i-
important and problematic

00:19:48.336 --> 00:19:52.035
field, but there are tons of discussions,
including philosophical

00:19:52.096 --> 00:19:56.055
discussions we could have about AI and
if it's good that it's there or

00:19:56.116 --> 00:19:59.396
not, but they all don't matter. It
is there, it will stay there,

00:19:59.436 --> 00:20:02.416
these companies will continuing to develop
these models.

00:20:02.436 --> 00:20:05.456
And even if they wouldn't,
the technology is there.

00:20:05.495 --> 00:20:09.356
Bad actors will have access to their own
models in the

00:20:09.396 --> 00:20:13.136
future. It does not matter at all
if companies like Anthropic or

00:20:13.235 --> 00:20:17.055
OpenAI continue. The technology
is there and the

00:20:17.096 --> 00:20:20.456
problems with it are also there,
therefore, and they will stay

00:20:20.515 --> 00:20:24.495
here. That would be my argument.
This argument here doesn't make

00:20:24.515 --> 00:20:28.285
too much sense to me. And therefore,
definitely scary.

00:20:28.376 --> 00:20:31.696
A scary world also from a cybersecurity
perspective.

00:20:31.795 --> 00:20:35.636
As I mentioned, I only see that, uh,
becoming worse

00:20:35.755 --> 00:20:39.076
in the future, and therefore,
maybe a career in

00:20:39.176 --> 00:20:42.815
cybersecurity is worth a second look
because yeah, that will

00:20:42.856 --> 00:20:44.156
be important.