Hacker News Reader

Comment by abidlabs | Hacker News Reader

Interesting to see Anthropic now downplaying the new vulnerabilities that Mythos discovered:

> We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass

View on HN

I think what they're saying is that this prompt/jailbreak only lets Mythos discover some really easy vulnerabilities that it probably fixes from a simple "Find and fix bugs in this code" and that this can be easily done by other models like GPT-5.5. Which is very different from targeted security research.

But it’s not that different from the whole premise of their red team scaremongering which was “we pointed the model at a source file and told it to find an exploit.”

> Interesting to see Anthropic now downplaying the new vulnerabilities that Mythos discovered:

That is absolutely NOT what is being said there.

They are referring to a very specific thing that you must have clearly seen and chosen to ignore—a jailbreak for LLMs that is used on other models and to some effect with Fable 5.

Dude, have people not realized there are a lot of anti-anthropic propaganda every since OpenAI started losing. It's all over reddit and twitter. So many bots.

Yeah cause Anthropic definitely doesn't do their own guerilla marketing all over the place. The entire internet is full of bots shilling all these garbage companies and their dogshit products.

Their hubris just became lethal for their business. Whoops, I guess.

AFAICT this is not talking about Glasswing stuff. They are saying that they were sent a demonstration of Fable 5 being used/abused in some specific way that led to the "discovery" of some minor, already-known vuln, and that other models can find it too. IOW, they're claiming that the USG's complaint is baseless and dumb.

Reply on news.ycombinator.com