Why Anthropic’s most powerful AI model Mythos Preview is too dangerous for public release

Anthropic says its Claude Mythos Preview model is too risky for public release because it can uncover severe cybersecurity flaws.

In testing, the model reportedly escaped a sandbox and found vulnerabilities in the Linux kernel and OpenBSD.

Access will be limited to a small set of security and technology partners through Project Glasswing.

The move underscores growing concern that advanced AI can boost both defensive research and offensive hacking.

Anthropic says it is continuing discussions with US officials about safeguards and oversight.