Anthropic accidentally leaked its most powerful AI model

A misconfiguration at Anthropic on March 26, 2026 exposed details about the company's most advanced AI model, internally known as "Claude Mythos" or "Capybara." Around 3,000 unpublished internal documents were inadvertently made publicly accessible through a content management system error.

Fortune was first to report the leak, after which Anthropic confirmed the model's existence. The company describes Mythos as a "step change" in AI capabilities, with dramatically higher scores on tests for software coding, academic reasoning, and cybersecurity compared to Claude Opus.

Leaked documents suggest that Anthropic has already completed training and is testing the model with select customers, particularly in cybersecurity applications. The release will proceed cautiously, as the company is concerned about the model's advanced capabilities and potential risks — including what internal documents describe as "unprecedented cybersecurity risks."

The documents also note that Mythos will be expensive to serve, with pricing reflecting this. The model is described as "larger and more intelligent" than existing Opus models. Anthropic confirmed the leak was due to human error in their content management system.

The incident highlights a growing challenge in AI: the more powerful the models become, the more critical strict security protocols are — not just for users, but for the companies building them.