Today, Chaofan Shou discovered that the npm package @anthropic-ai/claude-code version 2.1.88 contained a 59.8 MB source map file - essentially the complete source code of the project. Within a few hours, 512,000 lines of TypeScript code spread across GitHub (41,500+ forks).
Anthropic confirmed: "This is a release packaging error caused by human factors, not a hack."
What was found inside:
- 44 feature flags with hidden features that are already written but not enabled
- KAIROS - an autonomous agent mode that runs 24/7 in the background, subscribes to GitHub webhooks, and can "sleep" and resume work by itself
- BUDDY - a built-in tamagotchi with 18 types of creatures (duck, dragon, axolotl, capybara), a rarity system, and RPG stats like CHAOS, SNARK, and DEBUGGING
- Voice mode - full voice control via push-to-talk
- Browser control via Playwright
- Multi-agent mode - one Claude launches several working Claudes with limited toolsets
Unexpectedly, the code also contained an "undercover mode" that removes all traces of AI-written code when Anthropic employees commit to open source projects. They also discovered fake tools - decoy tools for data poisoning if someone records API traffic to train a competing model. And regex for detecting user frustration through profanity - in a company that makes LLM.
Technically - 1,900 TypeScript files, runs on Bun (not Node.js), terminal UI on React + Ink, 40 built-in tools, 50 slash commands.
