A group of users leaked Anthropic's AI model Mythos by reportedly guessing where it was located

Anthropic's Mythos model, which the company billed as too dangerous to release, has reportedly been accessed by an unauthorized third party, according to Bloomberg. The incident occurred on the day of its public announcement when a handful of users in a private Discord chat gained access. One member of the group, a third-party contractor for Anthropic, helped the others guess where the model was stored. This guess was based on previously leaked knowledge from another group, which hackers obtained from AI training startup Mercor. The group has not used the model for cyberattacks but has been running it continuously since its release and still retains access.

The leak raises significant concerns about the future of cybersecurity, particularly for frontier AI models. Anthropic had positioned Mythos as a high-risk model due to its advanced capabilities, yet the breach exploited contractor access and leaked operational knowledge. The incident highlights vulnerabilities in AI deployment security, where insider threats and fragmented information can bypass even stringent safeguards. As AI models grow more powerful, such breaches could lead to misuse, making robust access controls and contractor vetting critical for companies like Anthropic.