Anthropic Investigates Unauthorized Access to Claude Mythos Preview

The Story

Unauthorized users reportedly accessed Anthropic’s Claude Mythos Preview through a third-party vendor environment in early April. The model is designed for advanced cybersecurity tasks, with coding abilities Anthropic says can “surpass all but the most skilled humans” at finding and exploiting software vulnerabilities. Anthropic says it currently has no evidence that the reported access affected its own systems, and there is no current suggestion that malicious actors obtained the model.

The Facts

• Claude Mythos Preview is not a public model.

  Anthropic has no current plans to release it publicly because of concerns that it could be weaponized.

• The model is built for high-risk cybersecurity work.

  Anthropic says Mythos can identify and exploit vulnerabilities across major operating systems and web browsers when instructed to do so, with coding ability that can exceed all but the most skilled humans in this area.

• Official access is limited to selected organizations.

  Anthropic is making the model available through Project Glasswing to a limited group that includes Apple, Amazon, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, Nvidia, and other cleared organizations. It has also released Mythos to some technology and financial companies to help them secure systems against its reported vulnerability-exploitation capabilities.

• The alleged unauthorized access to Claude Mythos began in early April.

  The access reportedly occurred on the same day Anthropic announced limited Mythos testing for selected users.

• The users were reportedly part of a Discord channel.

  The channel is described as focused on finding information about unreleased AI models. The group reportedly used Mythos regularly but not for hacking, partly to avoid detection.

• The group reportedly used knowledge from another breach.

  Members used information from the Mercor data breach to make guesses about where the model could be found.

• The same group may have accessed other unreleased models.

  An individual connected to the reported access told Bloomberg that the group also has access to other unreleased models.

• Anthropic says the investigation is ongoing.

  The company is investigating a report of unauthorized access through one of its third-party vendor environments and says there is no indication the activity affected Anthropic systems.

Risks / Limitations

The main risk is not ordinary chatbot misuse. Mythos is described as a model capable of helping users find and exploit software vulnerabilities, which makes unauthorized access to Claude Mythos materially different from a leak involving a general consumer chatbot.

At the same time, Anthropic says it has no evidence that the reported access affected its own systems. There is also no current suggestion that malicious actors obtained the model, and the reported users allegedly avoided using it for hacking.

Background

Claude Mythos Preview was released for limited testing through Project Glasswing, a controlled access initiative involving major technology companies and other cleared organizations. The model is attracting government interest because of its cybersecurity capabilities: Inc. reports that the Commerce Department’s Center for AI Standards and Innovation and the National Security Agency already have access, while the Department of Treasury has sought access. The U.S. Cybersecurity and Infrastructure Security Agency reportedly does not have access.

What to Watch Next

The key question is whether Anthropic confirms the scope of unauthorized access to Claude Mythos, how the third-party vendor environment was exposed, whether other unreleased models were accessed by the same group, and whether vendor controls change for restricted AI systems.

Why This Matters

This case highlights a growing AI security problem: the most sensitive models are not only risky because of what they can do, but also because of who can reach them through contractors, vendors, and hidden infrastructure paths. As frontier AI systems become more capable in cybersecurity, access control may become as important as model safety testing itself.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.