Anthropic's 'Hacking Monster' Mythos Suffers Security Breach

Unauthorized Access to Preview Version Detected Concerns Grow Over Potential Misuse

The preview version of Anthropic's new artificial intelligence (AI) model "Claude Mythos" has been accessed without authorization.

Bloomberg reported Tuesday (local time) that a small number of unauthorized users gained access to Mythos. Anthropic immediately launched an investigation. "This technology is so powerful it could be exploited for dangerous cyberattacks," the company said.

Mythos was unveiled on the 7th of this month but has not yet been officially released. The preview version was provided to only a small number of trusted companies through a private program called "Project Glasswing." Anthropic explained that "so far, no access outside the 'vendor environment' has been detected." The vendor environment is a space where Anthropic's partners access the system for model development. Signs were also detected that source code Anthropic published on the open-source platform GitHub had been utilized. Security experts have warned that if Claude Mythos falls into the hands of hackers, attackers could exploit bugs faster than the defending organization can patch them.