A leak reveals that Anthropic is testing a more capable AI model "Claude Mythos"

https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/

by Tiberium • 4 months ago

66 24 4 months ago

fortune.com

dtagames• 4 months ago

https://archive.ph/ZMsZ1

saintfire• 4 months ago

The painful irony of bragging about/lamenting your new model's cybersecurity capabilities within and in response to leaking all the information about it due to poor cybersecurity.

Can't wait for the "$LAST_MODEL was amazing but this is the one that will change everything."

fakedang2• 4 months ago

The AI lab left the material, including what appeared to be a draft blog post announcing a new model, in an unsecured, public data lake

My tinfoil theory is that it was left by them to be discovered by the public.

moebrowne• 4 months ago

I believe a lot of 'leaks' you see these days are, at least some what, intentional.

charles_f• 4 months ago

The irony of bragging about how dangerous to cybersecurity it is with all the holes punches by the current generations

warwickmcintosh• 4 months ago

The "every new model is THE one" cycle is getting a bit old but the Capybara tier thing is actually worth paying attention to.

ggsp• 4 months ago

TFA states "Capybara and Mythos appear to refer to the same underlying model"

xvector• 4 months ago

Big if true. Any alleged "step change" over 4.6 is worth paying attention to.

Also, Hegseth did a big fucky wucky if true - since they went scorched earth with Anthropic, this could mean denying the US frontier model capabilities.

It would be pretty amusing if Anthropic emerges as the clear winner (for some period of time) and other governments can use them but the US government cannot.

Current admin is certainly too prideful to walk something like this back. After all, admitting mistakes is unmasculine, and we have very manly men in charge of war round these parts!

smackeyacky• 4 months ago

This is exactly the outcome you would expect authoritarians to generate. Not interested in facts, more interested in subservience. The good news is that surrounding yourself with incompetence is how authoritarian regimes fall. The bad news is, well, the country where this is playing out.

solenoid0937• 4 months ago

would be hilarious if they get declared a SCR and end up with the best model

enoint• 4 months ago

Particularly, if lax CMS leaked internal memos that their new model automatically uses vulnerabilities in CMS. And that’s what gets them attention as too risky.

gdevillers• 4 months ago

That model name does not look good in French: https://fr.wikipedia.org/wiki/Mythomanie ( https://en.wikipedia.org/wiki/Pathological_lying ).

pu_pe• 4 months ago

The English word for it is similar too (Mythomaniac)

gs17• 4 months ago

I think most English speakers who recognize it will think of it as meaning a collection of myths, as in the "Cthulhu Mythos".

futurecat• 4 months ago

absolutely. Also doesn't help that the slang for Mythomanie is the short form "mytho".

aiwlwlttok• 4 months ago

it's greek you peasants

Kye• 4 months ago

Yeah I'm confused by the comments here.

"a pattern of beliefs expressing often symbolically the characteristic or prevalent attitudes in a group or culture" (Merriam-Webster)

This doesn't seem obscure to me. It's what a model encodes in its weights. Am I falling into the xkcd 2501 trap?

fakedang2• 4 months ago

Why is it that in every thread of a newly released product, there's always some European popping by, just to say, "Oh, but the name doesn't sound appropriate in _my language_"? Suck it up Yurop, or make your own.

And imo, Mythos is a much better name than the kind of shit Mistral seems to come up with.

sdeframond• 4 months ago

Heehee, I'm still laughing about the Audi e-tron ^^

1attice• 4 months ago

Generalizing, generations of Europeans have been trained to hunt for these mistakes, in both their own writing and thinking and elsewhere, as they telegraph ignorance and open markets to competition.

It is standard in multilingual regions to consider multiple languages during product naming; to do otherwise is often considered humiliating.

You can bet your nut that "Mistral" was sounded out in more languages than "Cat, I farted" ever was.

Here are a couple of particularly famous fsckups:

- the Rolls Royce Silver Mist didn't sell well in Germany for some reason (coughs)

- the Chevy Nova didn't sell well in Spanish countries, again, seemingly for no reason (cackles)

CPLX• 4 months ago

This pattern of AI companies describing their own products as so spectacularly effective that they're dangerous really is a remarkable piece of propaganda engineering.

What is happening here would be easily understood and obvious by everybody if it the head of marketing for a food company was on TV talking about how pretty soon everybody will be eating their food, and how it's so unbelievably tasty that it might cause people to leave their families and abandon all other hobbies in pursuit of their delicious product, utterly destroying society as we know it.

I mean maybe it'll happen eventually. Maybe we'll all end up with a wire stuck in the back of our heads, floating in a vat of nutritious goo. But what we've seen so far has been an excellent, highly useful, and certainly groundbreaking industrial automation product.

Maybe they could just write a blog post telling us what the thing does and how much it costs and when we can try it.

unethical_ban• 4 months ago

Considering the impact this has had on cybersecurity stocks and the fears behind it, I'm shocked this isn't top news on HN.

blini-kot• 4 months ago

"accidental" data "leak" am i right

A leak reveals that Anthropic is testing a more capable AI model "Claude Mythos"

Related Stories