Anthropic’s rough week: leaked models, exposed source code, and a botched GitHub takedown

Anthropic accidentally leaked development files for a new model called Mythos and exposed a 59.8MB source map for Claude Code version 2.1.88.

Security researcher Chaofan Shou found the source map, revealing about 512,000 lines of Claude Code and full architecture details.

Anthropic requested GitHub takedowns under U.S. copyright law, which removed around 8,000 repositories before the company retracted broader notices.

An unsecured data store also exposed assets for Claude Mythos and details of a new Capybara tier, which Anthropic says is more capable than Opus.

Anthropic limited early access to Capybara to study cybersecurity risks amid concerns that exposed code could help attackers bypass protections.