Claude's leaked source code is absolutely UNHINGED!
Anthropic accidentally leaked Claude Code's complete source code twice in 13 months due to the same settings file error, revealing hidden features like 'undercover mode' that conceals AI authorship in open-source contributions. The leak exposed 512,000 lines of mostly AI-written code, unreleased model names, and features suggesting Anthropic is building autonomous AI workforce capabilities.
Summary
On March 31st, 2026, Anthropic made their second major source code leak in 13 months, accidentally exposing 512,000 lines of Claude Code's complete source code due to a missing line in a settings file. The leak occurred when a source map file was inadvertently included in a routine update, providing a direct link to unprotected code on Anthropic's cloud storage. This revealed numerous hidden features and raised questions about the AI safety company's security practices.
The leaked code contained over 80 hidden features, including 'Chyros' - an always-on mode that works in the background while users sleep, filing bug reports and managing projects autonomously. Another feature called 'Ultra Plan' sends tasks to remote servers where a second Claude plans entire projects for up to 30 minutes. The code also revealed internal names for unreleased AI models like 'Capybara', 'Numbat', and references to 'Opus 4.7'.
Most controversially, the leak exposed 'undercover mode' - a 90-line feature that activates when Anthropic employees contribute to open-source projects, instructing Claude to hide all traces of AI involvement and internal information. This raised questions about transparency from a company that promotes AI safety. The code also included a 'poison pill' feature that sends fake data mixed with real features to competitors who might intercept traffic.
The leaked codebase revealed that 100% of contributions were written by Claude itself, according to the head of Claude Code. This AI-generated code showed characteristics optimized for AI workflow rather than human readability, including massive files with thousands of lines and extensive dependencies. The code quality included concerning elements like 'god objects' and files connected to hundreds of other components, suggesting architectural challenges in AI-generated software.
Key Insights
- The head of Claude Code stated that 100% of his contributions to Claude Code were written by Claude Code itself, revealing that the entire half-million line codebase was AI-generated
- Anthropic built 'undercover mode' that instructs Claude to strip all traces of AI involvement when employees contribute to open-source projects, with instructions stating 'you are operating undercover' and 'do not blow your cover'
- The same packaging mistake caused source code leaks twice in 13 months, with the root cause being one missing line in a settings file that should have prevented source maps from being included
- Claude Code uses a second, smaller Claude running in the background to compress long conversations by deciding what to remember and forget, without users ever seeing this process happen
- The leaked code revealed 'Chyros', an always-on agent designed to work while users sleep, filing bug reports, reviewing code changes, and checking Git every 5 minutes with its own memory management system
Topics
Full transcript available for MurmurCast members
Sign Up to Access