Anthropic accidentally exposed the complete internal architecture of its sophisticated Claude Code AI assistant due to a source map leak in an npm package release. The 512,000 lines of code spread instantly, with clean-room Python rewrites and decentralized repositories rendering DMCA takedown efforts ineffective. The leak revealed hidden features and prompted widespread archiving, demonstrating the permanence of information in decentralized systems.
A significant source code leak at **Anthropic** has made the inner workings of its Claude Code AI permanently public. The company confirmed the exposure was caused by human error in a release package, with “No sensitive customer data or credentials were involved or exposed.”
Researcher Chaofan Shou was among the first to spot the leaked 59.8MB JavaScript source map file and posted a download link. The file contained 512,000 lines of code across 1,900 files detailing the AI’s full architecture, including LLM orchestration and 44 hidden feature flags.
The exposed code revealed internal systems like Kairos, a memory-logging daemon, and Buddy, an unreleased Tamagotchi-style AI pet with planned teasers. It also included an “Undercover Mode” subsystem designed to prevent the AI from leaking internal project names, as detailed by leaker Kuberwastaken.
Anthropic issued DMCA takedowns against GitHub mirrors, but the code’s spread became uncontrollable. Developer Sigrid Jin created a rapid clean-room rewrite called claw-code, porting the core to Python and gaining 30,000 GitHub stars.
Gergely Orosz argued in a post on X that such a rewrite “violates no copyright & cannot be taken down!” The legal standing is further complicated if Claude Code was written by AI, which affects copyright claims.
Decentralized platforms provided permanence, as the account @gitlawb mirrored the code to Gitlawb with a message it would never be taken down. This demonstrates how decentralized infrastructure, akin to cryptocurrency networks, lacks a single point of failure for content removal.
