Anthropic withholds Mythos Preview model because its hacking is too powerful

Anthropic is rolling out its new Mythos Preview model only to a handpicked group of tech and cybersecurity companies due to concerns about its ability to find and exploit security flaws. This "extremely autonomous" model can find "tens of thousands of vulnerabilities" and write exploits, unlike previous models. Anthropic is refusing public release until safeguards control its dangerous capabilities, as it successfully reproduced vulnerabilities and created proof-of-concepts in 83.1% of cases during testing. The company is providing up to $100 million in usage credits to over 40 organizations, including members of Project Glasswing like Amazon Web Services and Apple, to use the model for defensive security work. Anthropic is also briefing the Cybersecurity and Infrastructure Security Agency and the Commerce Department on the model's potential risks and benefits, while anticipating similar models from other AI companies within 6 to 18 months.