Anthropic’s most dangerous AI model just fell into the wrong hands

Bloomberg reported that a small group of unauthorized users accessed Anthropic’s restricted Mythos Preview through a third-party contractor environment, aided by insider access and online sleuthing.

Anthropic says it is investigating and has not found evidence that the breach extended beyond the vendor environment into its own systems.

The model was intended for cybersecurity research, so the exposure raises concerns about misuse of a frontier AI tool.

The incident also highlights weaknesses in vendor security and access controls around restricted AI models.