Anthropic inadvertently leaked over 512,000 lines of internal Claude Code through a source map published to npm. The error occurred due to a misconfiguration during a routine release.
Researcher Chaofan Shou shared the discovery early Tuesday, triggering massive interest online. Though Anthropic removed the package, mirrors and decentralized copies had already spread globally.
The leak revealed key components like "Kairos," an AI daemon managing memory consolidation, and "Buddy," a gamified assistant with customizable traits set for a limited rollout in April.
Notably, the leaked code included "Undercover Mode"-a mechanism meant to mask internal project names when contributing publicly.
Attempts at DMCA takedowns largely failed. Developer Sigrid Jin released a Python-based clean-room recreation called claw-code, gaining rapid traction with over 30,000 GitHub stars.
Legal experts argue such rewrites skirt copyright issues by being derivative yet functionally distinct. Meanwhile, decentralized hosting ensures perpetual access regardless of corporate action.
The incident highlights vulnerabilities in software distribution pipelines and raises questions about intellectual property rights in AI-generated systems.