WEBVTT

00:00:00.180 --> 00:00:04.140
As a software developer,
and I guess in general (laughs) as a

00:00:04.220 --> 00:00:08.200
human,
but especially as a software developer,

00:00:08.300 --> 00:00:10.940
Anthropic right now, if you want to
or not.

00:00:10.950 --> 00:00:14.890
And I don't think you should try to ignore
it because it is

00:00:14.920 --> 00:00:18.750
relevant. It's relevant for our future as

00:00:18.780 --> 00:00:20.500
software developers, I would say.

00:00:20.620 --> 00:00:24.580
And in this episode,
I won't talk about the Claude Code leak we

00:00:24.640 --> 00:00:25.660
had last week.

00:00:26.600 --> 00:00:29.800
I won't talk about their, uh,

00:00:29.940 --> 00:00:33.900
reinforced terms regarding the usage of
their subscription

00:00:33.960 --> 00:00:37.650
offerings, Claude Max and so on,
and how they're

00:00:37.740 --> 00:00:41.560
cracking down on unauthorized use of those

00:00:41.600 --> 00:00:45.580
subscriptions. They are doing
that right now because, of course, their

00:00:45.620 --> 00:00:49.220
subscription offerings,
just like the ones by OpenAI, are heavily

00:00:49.420 --> 00:00:53.340
subsidized and they can't make any money
if everybody

00:00:53.480 --> 00:00:57.000
maxes out their subscriptions. So yeah,
they're really

00:00:57.100 --> 00:01:00.740
restricting,
or they're trying to restrict,

00:01:00.750 --> 00:01:04.580
subscriptions to just humans,
just on their

00:01:04.640 --> 00:01:08.550
website, or in Claude Code
or the Claude desktop app, I

00:01:08.640 --> 00:01:12.530
guess. Uh, but yeah, again, this
is not the focus here and I

00:01:12.560 --> 00:01:16.200
won't even focus on their impressive
revenue

00:01:16.320 --> 00:01:19.950
growth, which is worth, uh, a,
a short note though, because

00:01:20.020 --> 00:01:23.660
Anthropic has reached a annual recurring

00:01:23.880 --> 00:01:27.530
revenue of $30 billion, which is

00:01:27.580 --> 00:01:31.240
already impressive,
but especially impressive

00:01:31.680 --> 00:01:35.320
$9 billion at the end of 2025. So, they

00:01:35.860 --> 00:01:39.640
more than tripled their annual recurring
revenue within

00:01:39.740 --> 00:01:42.580
just a few months, which
is really impressive.

00:01:42.640 --> 00:01:46.150
And therefore, of course,
if you wanna learn how to efficiently

00:01:46.640 --> 00:01:49.960
use Claude Code,
how to get the most out of it,

00:01:50.370 --> 00:01:54.210
And, um, it's also very popular,
which of course makes me happy,

00:01:54.280 --> 00:01:58.100
and you'll find a link below
if you wanna join it and learn how to

00:01:58.140 --> 00:02:00.300
efficiently work with Claude Code.

00:02:00.360 --> 00:02:03.800
But as mentioned,
that's not even the main topic here.

00:02:03.880 --> 00:02:07.700
Instead,
I wanna talk about Project Glasswing and

00:02:07.740 --> 00:02:11.480
their new model, Mythos,
which they haven't

00:02:11.560 --> 00:02:15.320
released to the public,
and they also shared

00:02:15.400 --> 00:02:19.020
why. And I think this
is important to understand, and it's also

00:02:19.040 --> 00:02:22.900
important to try to look behind the
scenes, behind their

00:02:22.920 --> 00:02:26.780
rationale,
and what the impact of this new model

00:02:26.880 --> 00:02:30.460
and how it works and its capabilities
is for us developers.

00:02:30.530 --> 00:02:33.890
So, what is Project Glasswing? What
is their new model all about?

00:02:34.580 --> 00:02:38.040
Uh, below, of course,
you'll find a link to this article as

00:02:38.140 --> 00:02:41.880
article on the official Anthropic site
where they announce Project Glasswing

00:02:42.000 --> 00:02:44.800
and also talk about their new model.

00:02:44.840 --> 00:02:48.400
And if I scroll down a bit,
we can already see some, uh,

00:02:48.600 --> 00:02:52.500
summary benchmark statistics here,
where we can see

00:02:52.540 --> 00:02:56.470
that this new model, the Mythos, uh,
preview version of the model, the,

00:02:56.580 --> 00:03:00.260
the, the model name is Mythos, um,
performs way better than

00:03:00.340 --> 00:03:04.320
Opus 4.6. And depending on which

00:03:04.380 --> 00:03:08.320
exact benchmark you look on, there
is quite a big difference

00:03:08.360 --> 00:03:11.660
between Opus 4.6 and this new model.

00:03:12.600 --> 00:03:16.130
Now, of course, this is, uh, on its own,
not super impressive.

00:03:16.220 --> 00:03:20.140
Whenever a new model is announced,
no matter by which company, it

00:03:20.220 --> 00:03:24.100
performs way better,
or at least a tiny bit better,

00:03:24.160 --> 00:03:26.180
models,
otherwise it wouldn't get released.

00:03:26.200 --> 00:03:29.860
And of course, there
are ways of gaming some of these

00:03:29.940 --> 00:03:33.600
so I typically don't care too much about
those benchmark numbers,

00:03:33.660 --> 00:03:37.520
and that wouldn't really be different for
this

00:03:37.560 --> 00:03:41.400
model here, but there
are interesting things about the

00:03:41.440 --> 00:03:45.180
new Mythos model. And that's the fact
that Anthropic

00:03:45.220 --> 00:03:49.100
decided to not release it to the public
because, as

00:03:49.180 --> 00:03:49.880
they say,

00:03:51.060 --> 00:03:54.280
it is too good at finding and

00:03:54.360 --> 00:03:58.340
exploiting vulnerabilities in operating

00:03:58.420 --> 00:04:02.320
systems, any other software, browsers,
really, in

00:04:02.680 --> 00:04:06.220
software in general. And in this article,
and also in a

00:04:06.400 --> 00:04:08.440
separate article, which
is also linked below,

00:04:09.240 --> 00:04:12.680
they share some details,
and especially this separate article is

00:04:12.740 --> 00:04:16.399
extremely long and gives concrete

00:04:16.459 --> 00:04:19.740
examples for vulnerabilities and potential

00:04:19.760 --> 00:04:23.420
exploits this new model found.

00:04:24.200 --> 00:04:27.740
Uh, for example,
they start in this article with a

00:04:27.780 --> 00:04:31.050
very serious exploit and vulnerability

00:04:31.620 --> 00:04:34.560
that was found in, uh,

00:04:34.640 --> 00:04:38.600
OpenBSD. OpenBSD is, of course,
an operating system which is

00:04:38.640 --> 00:04:42.040
popular on certain pieces of networking
software, for example,

00:04:42.540 --> 00:04:45.980
and Mythos, their new model running in

00:04:46.060 --> 00:04:49.860
Agentic Harness, like Claude Code,
I guess, was able

00:04:49.940 --> 00:04:53.600
to find and exploit, a- and that's the

00:04:53.620 --> 00:04:57.170
interesting part - a vulnerability
that was related

00:04:57.220 --> 00:05:00.860
to integer overflow and, uh, memory

00:05:00.940 --> 00:05:04.680
access, unexpected memory access, um,

00:05:05.040 --> 00:05:08.400
that was able to crash machines that
were running

00:05:08.460 --> 00:05:12.400
OpenBSD in a reproducible way, which

00:05:12.410 --> 00:05:16.400
could, of course,
be leveraged to run very damaging,

00:05:16.520 --> 00:05:19.940
uh, denial of service attacks by over
and over

00:05:20.000 --> 00:05:23.800
sending specific packets
and requests to such machines

00:05:23.820 --> 00:05:27.400
here, which exploited
that vulnerability to bring

00:05:27.440 --> 00:05:31.280
down those machines
and potentially bring down entire

00:05:31.320 --> 00:05:34.980
corporate networks.
And this vulnerability was

00:05:35.000 --> 00:05:38.720
detected in a run that cost under $50,
though

00:05:39.340 --> 00:05:42.600
the overall set of runs cost under

00:05:42.620 --> 00:05:46.360
$20,000. And since you, of course,
don't know in advance which

00:05:46.500 --> 00:05:50.160
run will find a vulnerability, it's
that number that matters.

00:05:51.240 --> 00:05:55.110
Still, of course,
it's easy to imagine that

00:05:55.180 --> 00:05:58.840
a model capable of finding such critical

00:05:58.880 --> 00:06:02.280
vulnerabilities for such a comparatively
low

00:06:02.360 --> 00:06:06.310
cost...Depending on who you are,
if you're a nation, um, for

00:06:06.380 --> 00:06:08.270
example, or some, uh,

00:06:09.120 --> 00:06:13.000
serious bad actor,
that may not be a lot of money

00:06:13.060 --> 00:06:13.400
to you.

00:06:14.210 --> 00:06:17.540
That, of course,
is a problem because it's easy to imagine

00:06:17.580 --> 00:06:21.040
that if such a model were

00:06:21.060 --> 00:06:24.740
developed by a company, an organization

00:06:25.660 --> 00:06:28.740
that cares a bit less about security

00:06:29.200 --> 00:06:32.510
and/or that maybe doesn't have to

00:06:32.640 --> 00:06:35.900
fear any consequences of

00:06:35.980 --> 00:06:39.060
abusing such vulnerabilities,

00:06:40.160 --> 00:06:43.880
that this could be a problem. And it seems

00:06:44.340 --> 00:06:48.240
as if we're entering a new age

00:06:48.640 --> 00:06:51.200
with AI, with these AI models,

00:06:52.920 --> 00:06:56.760
where nothing is secure and

00:06:56.820 --> 00:07:00.180
it's easier than ever to mass

00:07:00.200 --> 00:07:03.940
deploy AI agents running models like this
to

00:07:04.040 --> 00:07:07.630
scan all kinds of pieces of software
and find and

00:07:07.640 --> 00:07:09.640
potentially exploit vulnerabilities.

00:07:09.680 --> 00:07:13.390
And of course, as a human on, on your own,

00:07:13.540 --> 00:07:15.960
there is no way of keeping up with it.

00:07:16.040 --> 00:07:19.660
I mean, the bug, the exploit that
was found here existed

00:07:19.740 --> 00:07:23.700
for, I think they said 27 years
or something like this.

00:07:25.040 --> 00:07:28.620
This shows that no human was able to

00:07:28.700 --> 00:07:32.580
find this bug in such a long period of
time, including

00:07:32.620 --> 00:07:36.380
bad actors,
which of course would've had an interest

00:07:36.390 --> 00:07:39.400
attack this operating system in the past
too.

00:07:39.800 --> 00:07:43.500
Now this is just one,
maybe the most prominent

00:07:43.580 --> 00:07:47.380
finding, um, this new model had. They
are listing

00:07:47.600 --> 00:07:51.500
way more bugs and exploits the model
was find and

00:07:51.580 --> 00:07:53.460
sometimes also able to abuse,

00:07:54.500 --> 00:07:56.320
and they also shared other stories.

00:07:56.600 --> 00:07:59.920
Uh, on X, for example,
like the model being able to

00:07:59.980 --> 00:08:03.950
escape a sandbox or the AI agent running
the

00:08:04.040 --> 00:08:07.420
model being able to escape a sandbox in
which it was running.

00:08:08.300 --> 00:08:11.979
And that brings us back to Project
Glasswing, which is an

00:08:12.040 --> 00:08:15.980
initiative created by Anthropic together

00:08:16.140 --> 00:08:19.260
with other big companies like AWS, Apple,

00:08:19.920 --> 00:08:23.480
um, Microsoft, the Linux Foundation,
and others

00:08:23.620 --> 00:08:27.200
to use this model to basically patch

00:08:27.300 --> 00:08:31.190
up their software before this is

00:08:31.220 --> 00:08:34.979
publicly released
and before the public gets their hands

00:08:35.050 --> 00:08:38.900
on this model. That is the,
the narrative of this article.

00:08:39.020 --> 00:08:41.320
That is the explanation of Anthropic.

00:08:42.840 --> 00:08:43.229
And

00:08:44.140 --> 00:08:47.100
I have some mixed thoughts here. Now,
for one,

00:08:48.320 --> 00:08:52.040
I have no strong reason to believe
that this isn't true.

00:08:52.420 --> 00:08:56.000
Clearly, Anthropic would have some

00:08:56.180 --> 00:09:00.080
reasons to not release this model outside
of what they're

00:09:00.140 --> 00:09:01.520
mentioning here. For example,

00:09:02.340 --> 00:09:06.020
um, I read that this model is a roughly

00:09:06.160 --> 00:09:10.140
10 trillion parameters model, which
is way bigger than

00:09:10.240 --> 00:09:14.120
all the frontier models we had thus far
or were able

00:09:14.160 --> 00:09:16.080
to use publicly thus far,

00:09:16.960 --> 00:09:20.640
and training it is said to have cost
around

00:09:20.660 --> 00:09:24.560
$10 billion. The token cost of

00:09:24.620 --> 00:09:28.380
this model, I read, is expected to

00:09:28.420 --> 00:09:31.380
be around this range, $25,

00:09:31.430 --> 00:09:35.300
$125 for, uh, input and output tokens.

00:09:35.580 --> 00:09:39.320
And of course,
that would also be reasons for not

00:09:39.360 --> 00:09:43.220
releasing this model because they can't
include it in

00:09:43.240 --> 00:09:46.670
their Claude subscriptions because it's
just too

00:09:46.700 --> 00:09:50.620
expensive.
They would have to ramp up the

00:09:50.660 --> 00:09:53.550
to a price point where not a lot of people
are willing to pay it,

00:09:54.810 --> 00:09:58.420
uh,
and therefore there wouldn't really be a

00:09:58.460 --> 00:10:01.640
exposing it to the public,
at least as part of Claude code.

00:10:01.680 --> 00:10:05.449
Now of course,
they could still expose it through their

00:10:05.480 --> 00:10:09.400
pay-per-use cost basis, and
if it's expensive, who

00:10:09.460 --> 00:10:13.240
cares? If there are companies, people
that would be willing to pay it,

00:10:13.460 --> 00:10:16.740
they could do that. And that, of course,
is the part where the

00:10:16.780 --> 00:10:20.020
cybersecurity concerns might really come
into play

00:10:20.079 --> 00:10:23.880
because clearly that all

00:10:23.940 --> 00:10:27.400
is very likely not made up. I mean,
it's definitely not made

00:10:27.480 --> 00:10:31.220
up. The FFmpeg team, for example, which

00:10:31.300 --> 00:10:35.160
is also listed here as a, um, as a

00:10:35.220 --> 00:10:38.900
vulnerability that they were f-
able to find a vulnerability in

00:10:38.980 --> 00:10:39.740
FFmpeg,

00:10:40.600 --> 00:10:44.000
uh, the team confirmed on X that Anthropic

00:10:44.180 --> 00:10:48.000
sent, uh, sent a patch for, uh, uh, a

00:10:48.040 --> 00:10:51.050
vulnerability in the FFmpeg, uh,

00:10:51.300 --> 00:10:54.960
software, uh, program So yeah, this is

00:10:54.980 --> 00:10:58.600
clearly not made up. These concerns
are valid.

00:10:59.080 --> 00:11:02.500
Cybersecurity concerns are valid,
especially since, of course,

00:11:03.060 --> 00:11:06.780
if money is not the main issue,
you could deploy thousands of

00:11:06.980 --> 00:11:10.820
agents running simultaneously using this
or

00:11:10.940 --> 00:11:14.660
similar models,
which we may have in the future to, yeah,

00:11:15.000 --> 00:11:18.520
scan all kinds of software
and exploit them.

00:11:18.560 --> 00:11:21.000
And of course, the big problem is that

00:11:22.560 --> 00:11:26.400
using this model to find vulnerabilities

00:11:26.440 --> 00:11:29.350
and patch them up is possible,

00:11:30.180 --> 00:11:33.780
but it's only possible if the

00:11:33.880 --> 00:11:37.600
owner or the maintainer of a certain piece
of software can afford the

00:11:37.660 --> 00:11:40.900
model or gets access for free
or anything like that.

00:11:41.240 --> 00:11:45.020
And even if a vulnerability is patched up,

00:11:45.180 --> 00:11:45.900
we all know

00:11:46.860 --> 00:11:50.760
that not all (laughs) computers out there,
not all machines,

00:11:50.860 --> 00:11:52.080
not all users

00:11:53.140 --> 00:11:57.000
have up-to-date software running on them.

00:11:57.120 --> 00:12:00.260
If you were to take a look at all the

00:12:00.280 --> 00:12:03.640
various servers running out there in the

00:12:03.680 --> 00:12:07.380
worldwide webI would guess the vast

00:12:07.420 --> 00:12:10.660
majority of them is running outdated
software.

00:12:10.760 --> 00:12:14.460
I mean, on our phones, our, our laptops,
we are

00:12:14.900 --> 00:12:18.460
of- not running the latest software,
the latest

00:12:18.860 --> 00:12:22.760
version of our operating system.
The latest security patch may not be

00:12:22.780 --> 00:12:26.440
installed and that
is true for all layers of s- of

00:12:26.520 --> 00:12:30.460
software.
And in a world where it's easier than ever

00:12:30.580 --> 00:12:34.540
find security vulnerabilities,
that of course becomes an

00:12:34.640 --> 00:12:38.390
even bigger issue.
Because of course the good thing about

00:12:38.460 --> 00:12:41.880
this AI model is that it can also be

00:12:41.960 --> 00:12:45.800
used for proactively looking for security

00:12:45.900 --> 00:12:49.800
vulnerabilities and patching them.
So it's not just a tool for attackers.

00:12:49.840 --> 00:12:53.620
It can also make defense easier because
you now have a

00:12:53.680 --> 00:12:57.520
tool that can be run simultaneously in
parallel across

00:12:57.530 --> 00:13:00.380
thousands of agents to make your software
secure.

00:13:00.440 --> 00:13:04.380
In theory,
this can be a very useful tool for

00:13:04.480 --> 00:13:08.340
of course, again, not every company,
person

00:13:08.400 --> 00:13:12.260
that may be developing crucial software
may be able to afford it, may

00:13:12.340 --> 00:13:16.240
be interested in using it,
and then even if it is

00:13:16.320 --> 00:13:18.900
used to find and patch vulnerabilities,

00:13:19.540 --> 00:13:23.340
still these latest versions will not be
installed

00:13:23.400 --> 00:13:27.220
everywhere.
And that of course gives attackers, uh,

00:13:27.300 --> 00:13:31.030
opportunity where they know about way

00:13:31.120 --> 00:13:34.680
more vulnerabilities than before at some
point because

00:13:35.560 --> 00:13:39.490
way more vulnerabilities are detected,
but not every

00:13:39.520 --> 00:13:42.960
machine, every user
is protected against those

00:13:42.980 --> 00:13:46.730
vulnerabilities and that
is one real concern I have

00:13:46.820 --> 00:13:50.600
about this development. Now that is the,
the bigger

00:13:50.640 --> 00:13:54.480
picture which affects everybody,
all companies, all humans

00:13:54.560 --> 00:13:58.100
in the end. Another question of course
is what does

00:13:58.580 --> 00:14:02.000
a model like this mean for us

00:14:02.320 --> 00:14:05.760
developers? I mean,
clearly this seems to be a

00:14:05.940 --> 00:14:09.820
highly capable model that
was able to look for

00:14:10.040 --> 00:14:14.000
vulnerabilities on its own
and exploit vulnerabilities on its

00:14:14.060 --> 00:14:17.900
own. So yeah, what is the impact for

00:14:17.940 --> 00:14:18.700
developers?

00:14:19.820 --> 00:14:22.880
And I think here, when it comes to that,

00:14:24.180 --> 00:14:28.020
not a lot changes for now. I

00:14:28.080 --> 00:14:31.840
mean,
we're already living in a world where

00:14:31.920 --> 00:14:35.600
AI agents like Claude code
and the underlying models,

00:14:35.640 --> 00:14:39.530
true for Codex and so on,
whatever your favorite AI agent

00:14:40.280 --> 00:14:42.800
are able to generate most of our code.

00:14:43.200 --> 00:14:45.280
You may not be using them,
you may not like them.

00:14:45.600 --> 00:14:49.510
I created a separate video where I share
my feelings about that and that

00:14:49.580 --> 00:14:53.560
this sucks the joy out of the,
the software development part

00:14:53.600 --> 00:14:57.370
for me, but it is the,
the reality nonetheless, if you like it or

00:14:57.460 --> 00:15:01.280
not. And believe me,
I don't necessarily like it, but yeah,

00:15:01.340 --> 00:15:05.280
reality. Nonetheless,
what a human brings to the

00:15:05.320 --> 00:15:09.220
table or why humans still matter here
and may matter

00:15:09.420 --> 00:15:13.320
more than ever is, of course,
that you definitely

00:15:13.329 --> 00:15:17.200
don't want an AI agent like this go rogue

00:15:17.260 --> 00:15:20.580
and work totally on its own. Steering

00:15:21.000 --> 00:15:24.560
such models and agents, controlling them,

00:15:24.780 --> 00:15:28.540
giving them clear tasks,
limiting the scope of the

00:15:28.620 --> 00:15:29.540
work they do,

00:15:31.000 --> 00:15:34.340
all these things are more important than
ever.

00:15:34.549 --> 00:15:38.220
These models can, as it seems, do way

00:15:38.320 --> 00:15:39.140
more than

00:15:39.980 --> 00:15:43.820
the vast majority of developers can do.
Definitely way more than I can do.

00:15:45.420 --> 00:15:49.160
And yet,
when it comes to shipping products,

00:15:49.260 --> 00:15:52.700
comes to building software used by humans,

00:15:54.360 --> 00:15:57.960
the influence of a human is of

00:15:58.060 --> 00:16:00.500
course extremely, extremely important.

00:16:00.600 --> 00:16:04.300
What's changing, of course,
is our role as software developers.

00:16:04.400 --> 00:16:08.260
We're changing from the people that
are writing the code to the

00:16:08.320 --> 00:16:11.620
people that are steering the model,
that are reviewing the code, that are

00:16:11.720 --> 00:16:15.340
understanding what it does, that
are setting the scope.

00:16:15.400 --> 00:16:19.340
And yeah, again,
I talked about this in that other video,

00:16:19.400 --> 00:16:22.940
and then that this may not necessarily be
what, uh,

00:16:23.179 --> 00:16:26.920
what, what you like.
It's definitely not w- why I got

00:16:26.940 --> 00:16:28.800
into software development in the first
place.

00:16:28.880 --> 00:16:32.620
But yeah, this is the impact here
and the more capable these

00:16:32.660 --> 00:16:36.600
models get,
the more important I think it gets to

00:16:36.780 --> 00:16:40.510
have that,
that human voice in there as well,

00:16:40.600 --> 00:16:44.200
that human influence, uh, in there as well

00:16:45.280 --> 00:16:48.620
So that is, that is
that changing role and, and

00:16:49.320 --> 00:16:53.020
our role in the future. But yeah, I mean,

00:16:53.900 --> 00:16:57.160
these are really interesting developments
and

00:16:57.780 --> 00:17:01.680
especially this model
and its implications and

00:17:01.860 --> 00:17:05.130
that cybersecurity relevance it has

00:17:06.080 --> 00:17:09.360
makes one think what would have happened
or what would

00:17:09.420 --> 00:17:12.490
happen if other

00:17:12.960 --> 00:17:16.110
actors, other nations or

00:17:16.180 --> 00:17:19.920
organizations in the world get their hands

00:17:20.020 --> 00:17:23.760
on this model or models that
are similar in capability.

00:17:23.800 --> 00:17:27.060
Because of course it's only a matter of

00:17:27.140 --> 00:17:30.630
time until models with similar

00:17:30.660 --> 00:17:33.360
capabilities are accessible

00:17:34.240 --> 00:17:37.900
by the public or certainly at least by

00:17:38.840 --> 00:17:41.440
other nations and actors. And

00:17:42.380 --> 00:17:46.120
yeah, I'm not sure if we
are prepared for that

00:17:46.240 --> 00:17:50.220
new race in cybersecurity and that delay

00:17:50.260 --> 00:17:54.240
between bugs being found and being patched

00:17:54.420 --> 00:17:57.840
and people installing those patches.

00:17:58.160 --> 00:18:02.140
I think we'll enter a new era of
cybersecurity and we'll

00:18:02.200 --> 00:18:06.180
be able to adjust, I'm sure, but this, uh,
definitely

00:18:06.240 --> 00:18:07.710
marks an interesting,

00:18:08.600 --> 00:18:12.510
uh, point in the history of,
of model development

00:18:13.240 --> 00:18:13.640
I would say