350.6K
DAOn March 31, Anthropic accidentally exposed the source code for Claude Code through a debug (.map) file in their public npm package. This made it possible to reconstruct the original codebase, which was quickly identified and shared online.
Around 500,000 lines of code across ~1,900 files were accessible. This included the TypeScript source code for the CLI, along with internal tools, agent workflows, and system prompts that define how the system behaves.
Importantly, the core Claude model itself was not leaked. Model weights, training data, and backend systems were not part of this exposure.
People analyzing the code found references to internal models like “Capybara,” along with hidden features, agent tools, and details on how the system is structured and operates.
The issue was traced back to debug source maps being included in a production build, so this wasn’t a hack, just an accidental exposure during deployment.
#claudecode #claudecodeleak #anthropic #technews #ai
@dailysakshi
