Anthropic introduced the Mythos model, but access to it is restricted due to its outstanding capabilities in cyber attacks

Anthropic launches limited access to the new Mythos model

Anthropic announced the release of its latest AI model, Mythos, designed specifically for finding critical vulnerabilities and modeling cyber threats. According to Frontier Red Team leader Logan Graham, Mythos demonstrated such powerful hacking capabilities that it cannot be released publicly.

Who will have access
* A restricted list – about 50 companies and organizations working with critically important IT infrastructure.
* Early users include Amazon, Apple, Google, and the Linux Foundation (according to *The Wall Street Journal*).

Why the model is considered dangerous
Metric Mythos Previous model Claude Opus 4.6 Vulnerability search effectiveness ~10× higher Time required to find bugs instant vs. 2 weeks > 1 instant result Number of critical errors in Firefox (start of year) > 200-
New Stanford research shows that neural networks can find and exploit errors in real networks almost as quickly as professional hackers. This confirms experts’ concerns: “the window between discovering a vulnerability and creating working software to exploit it is shrinking” (additionally, *Mythos* can do this 10 times faster).

What Logan Graham says
> “We don’t plan to release Mythos publicly. But in a few years other AI systems will be able to have similar hacking capabilities.”

He urges the IT community to prepare for a new reality where the delay between identifying a vulnerability and exploiting it could become zero.

Conclusion
Anthropic limited access to Mythos only to selected corporations due to concerns about its hacking potential. The model has already shown impressive bug‑finding effectiveness, heightening experts’ worries about the rapid development of AI hacking and the need to prepare for future threats.