So, the AI commentator community seems to have

lost its ever-loving mind. Again.

This time, it's about a new announcement about

Anthropic's new ""Project Glasswing"" and the

underlying

Claude Mythos model that they say they're

not releasing because they say it's too

powerful.

Some people are saying this is incredibly

dangerous and it's very good that Anthropic

is holding off on the release, and other people

are saying this is just a publicity stunt.

Here's the reality:

Both of those things are true.

Well, to an extent. Kind of. We'll get into it.

One thing is, the model itself is probably the

least relevant part of the whole thing.

And the publicity stunt part? We've been seeing

a lot of this particular attention grabbing

technique lately.

I'm sure it's only going to be getting worse as

the AI companies get more desperate to

keep up the investment dollars flowing so they

can be shoveled into the fire.

How are we supposed to believe this Shhh----?

This is the Internet of Bugs.

My name is Carl.

I've been a software professional since the 1980s and I'm

trying to do my part to make the Internet

a safer and less buggy place.

You can find links where you can get in touch

with me at InternetofBugs.com if you're so

inclined.

I'm not going to spend a ton of time going over

what Claude Mythos actually is.

I'll put some links in the description if you

need to catch up.

The short version is: Anthropic announced that

they have this new model they're calling Mythos,

but they say it's too dangerous to release it

to the public.

It is so good at finding and exploiting bugs,

they say, that they "believe it could reshape

cybersecurity" so they created something they're

calling "Project Glasswing" in conjunction

with a bunch of big tech firms quote: "in an

effort to secure the world's most critical

software" unquote.

Let me give you the short version of what's

going on and then I'll give you the more

details.

So there are three things at play here.

First off, Anthropic is doing that "We're

telling you that this thing is so great that

we're not letting you see it, we're only

getting access to it to people we want to

and you're going to have to take our word and

their word for how great it is."

The second thing that's going on is a pattern I'm

starting to see more often where AI companies

spend a ton of money on something that's

legitimately useful, then they attribute the

useful result

of all that spending to the functionality of

their software and not the amount of money

they actually spent on it.

And the third thing is they're making a big

deal about the one particular kind of task

they have a good way of a model doing, in the

hope that the press and the public will think

"Well, if that new AI is so good at X, then it

must also be good at all this other stuff,

right?"

So if you want a one- sentence takeaway it's

this: "The security risks the news are reporting

are real, but they're not because of how good

the new AI model is, so much as they're really

about how much money the company is spending to

show off this one specific scenario that

they've invented to make their AI look better

than it probably actually is."

All right, so let's get into some details.

Bragging about an AI product that you haven't

actually released is a pretty common pattern.

OpenAI let a select few people have early

access to ChatGPT-5, and many of them reported

how fantastic it was! And then they had to walk

it back after it was released and people

with no incentive to make OpenAI happy started

really testing it objectively, and it landed

with a thud.

A similar thing happened with DEVIN the so-called

"first AI software engineer" where they released

some demo videos and quotes from people that

they had handpicked to show it to, and then

the demos turned out to have been greatly

exaggerated. And Devin has just utterly failed

to have the impact that they claim.

When companies do this, you have every reason to

be skeptical.

For example, in this case, if it was really as

dangerous as they're saying and they really

were as concerned about the safety implications

as they say, then a responsible company - unlike

Anthropic - would just shut the hell up about it until

all the bugs the new AI had found had

been fixed. That's called "responsible disclosure".

It happens all the time. Where security experts

find the vulnerability, they notify the affected

company about the problem, and then they wait

for the fix to be released before they announce

their involvement.

So, next item up is how AI companies have

started trying to get you to confuse effort

and money for model ability.

I made a video recently about how OpenAI had

run a custom-built internal model for hours

to simplify some math equations, got some

academics try to paper about it, and then wrote

a press

release that made it sound like ChatGPT was a

PhD physicist and had discovered new science.

This is the same kind of thing. Some security

folks over at aisle.com did some great research

and wrote up a great piece about it sub titled

"why the moat is the system and not the model."

You should check that out if you want more

detail - I've linked it below. They took some

of the more celebrated bugs that Anthropic says

the new model found and they tested them

on small, cheap, open-weight models, and got very

similar results. That's a very strong evidence

that the new model isn't the real story here. It's

not that spectacular. What's important

is the amount of time and money that they spent.

So here's an excerpt from an Anthropic blog

post about the denial of service bug they found

in OpenBSD which is a very secure operating

system that I've been running on my own servers

since the 1990s.

They say that the bug was found after a

thousand runs at a cost of around twenty

thousand dollars

likewise they spent ten thousand dollars to

find a bug in FFmpeg, and they say they

found a few thousand other bugs. Now they don't

tell us the total amount of money they spent

on the computer time for the project, but if we

were to extrapolate from the OpenBSD

and FFmpeg bug costs times a few thousand

for the other bugs, we could easily be at tens

or hundreds of millions of dollars. And consider

too that they almost certainly spent a lot of

money on computer cycles looking at software

that they didn't have any bugs that they found.

They also said they were gifting a hundred

million dollars of compute to their partner

companies in "Project Glasswing", and four

million dollars in grants to open source groups.

Let

me put that in perspective for you: The biggest

bug hunting program in the world, HackerOne,

spends something like eighty to ninety million

dollars a year total. "Project Glasswing" is

spending

125% of that just

in compute for their partner companies. They

could easily have spent an additional multiple

of that searching for the bugs that were

announced

in this press release, to say nothing of however

much was spent training in the model on

existing

bugs while it was being built in the first

place. Anyone who spends that much money looking

for bugs is going to find a ton of them, and we

should all be happy that they were willing

to spend that, and the end result of this will

be a much safer Internet for everyone, and

that's fantastic. And Anthropic should absolutely

be commended for that. But they're not saying

"we spent a ton of money making Internet safer

for you" they're saying "look how dangerous

and powerful our new model is - you should be in

AWE of it" and that's just not the reality

here. Lastly, they're glossing over that there is

a big difference between discrete tasks

with well-defined success criteria and the

ambiguities that we as humans deal with all

the time. It's very straightforward to set up an

environment with a checklist of steps

to use to see if a bug has been found, and then

let the AI run in that environment for

hours days or even weeks, until you either find

something or you decide you've spent enough

money and it's time to look elsewhere. But this

is much more like the way that an AI

is taught to be good at chess than it's

training for human equivalent general

intelligence.

And those are two different things, and getting

better at one doesn't help you with getting

better at the other. The AI Security Institute

in the UK has access to the model and they

ran it against a number of their "Capture the

Flag" scenarios that they use in hacking

competitions. It did pretty well - better than any

other AI they've tested, but keep in mind

that doing hacking Capture the Flag problems is

working with clear rules and well-defined

victory conditions. These are more like playing

chess or go than general software building.

So it's vastly easier to train an AI to do this.

So, what does that mean for you and for

the Internet? Well it means that, as has

happened several times before, when new bug

hunting

tools were released or new bug hunting

techniques were published, we're going to be

going through

a bumpy time for a while from a security

standpoint. Expect a lot of security updates for

your

phones, tablets and laptops, a lot of news

reports about hackers for a while, and once

we get through this remediation process for

everything that this initiative is discovering

we'll all hopefully be in a better place and

things will calm down. It also means that

- finally -

AI has given us something that might help

counteract all the nightmare security holes

that vibe coding is creating. Now I doubt that

most vibe coders are going to spend the time

and effort using this kind of technique to look

for the exploits in their code, but they

could, and that's more than we had before. It's

not a silver bullet though. Finding and

preventing

the customer-facing network and OS-related

kinds of things that Capture the Flag games

prioritize, like buffer overruns, or remote access,

privilege escalation, that kind of stuff - that's

great, but there are a lot of other kinds of

bugs that aren't so easy to look for, like

logic errors, data loss, data corruption,

synchronization errors, half-committed

transactions, roach

motel UIs, whack-a-mole, and that kind of

stuff. We're still a very long way from

a "true AI software engineer". So, don't panic, don't

let the news stress you out, but do try

to be extra vigilant for a while. Thanks for

watching. Let's be careful out there.