Claude Mythos Leak: Model 'Too Dangerous for Public' Falls Into Unauthorized Hands

Anthropic has found itself in an awkward situation: the Claude Mythos model, which the company had spent weeks presenting as too dangerous for open access, leaked to outsiders before official testing even began. According to Bloomberg, a small group of unauthorized users gained access to Mythos on the very day Anthropic announced plans to provide the model to a limited circle of companies. The very fact of Mythos's existence became known earlier — through a leak, not an official announcement.

The company has now stated it is investigating the incident. Mythos was positioned as a model with exceptional capabilities in the field of cybersecurity. These very capabilities became the official justification for an extremely cautious approach to the release: Anthropic insisted the model was not ready for a broad user base due to potential risks.

The company had planned to launch closed testing only with selected partners — as part of the Glasswing project. The incident puts Anthropic in a particularly delicate position. The company has built its reputation around the idea of responsible AI development and strict security controls.

To publicly declare a model too dangerous — and then allow its leak on the day of announcement — is a blow precisely to the part of Anthropic's image that the company has cultivated for years. It is currently unknown how users gained access, how complete that access was, and what capabilities of the model they managed to leverage. Anthropic is not disclosing the details of the investigation.

What happened exposes a systemic problem faced by all major AI laboratories: the more closed and "secret" a model is declared to be, the higher the interest in it — and the more painful any leak becomes. For Anthropic, this incident has become a tangible test of how much the company's words about safety are backed up by actual processes within the company.