Anthropic Accidentally Leaks 'Claude Mythos' -- a New Model Tier Above Opus with Unprecedented Cybersecurity Risks

TL;DR: A misconfigured content management system at Anthropic exposed nearly 3,000 unpublished assets, including a draft blog post revealing 'Claude Mythos' (also internally called 'Capybara') -- a new model tier sitting above Opus 4.6, described as a 'step change' in capability. Anthropic confirmed the leak, calling it a 'human error,' and acknowledged the model is being trialed with early access customers. The draft post explicitly flagged that Mythos poses unprecedented cybersecurity risks, including the ability to identify and exploit software vulnerabilities. Leaked documents also revealed a private CEO summit in Europe tied to enterprise sales.

Why it matters: Mythos/Capybara would be Anthropic's most powerful model yet -- a new tier beyond Opus -- arriving at a moment of intense frontier competition. The cybersecurity risk framing signals Anthropic is taking a cautious approach to release. The accidental exposure of internal roadmap materials also raises questions about operational security at top AI labs.

Source: Fortune