Tragic mistake... Anthropic leaks Claude’s source code
Fireship dissects Anthropic’s accidental Claude Code source leak, revealing anti-distillation traps, undercover mode, and hidden feature flags including a background agent called Chyus.
- Anthropic’s Claude Code npm v2.1.88 shipped a 57 MB source map containing 500,000+ lines of TypeScript, discovered by security researcher Chiao Fan Sha.
- Root cause likely Bun.js (recently acquired by Anthropic): a known open GitHub issue about Bun serving source maps in production.
- Claude Code uses anti-distillation poison pills — fake tool references designed to mislead competitors training models on Claude outputs.
- Undercover mode instructs Claude to never mention itself in commit messages or outputs, framed as preventing model-name leaks but widely read as covert AI laundering.
- A regex pattern-matcher scans prompts for profanity keywords to detect user frustration and log an event — no AI involved.
- Leaked feature flags reference Opus 4.7, a model called Capiara, ultra plan, coordinator mode, demon mode, and Chyus — a background agent with a daily journal and dream-mode memory consolidation.
- Python rewrite Claw Code (using OpenAI Codex) became the fastest repo to reach 50,000 GitHub stars; a separate fork called OpenClaw works with any model.
- The codebase is dense with comments interpreted as AI-readable instructions, suggesting Claude was partly writing its own coding tool.
2026-04-01 · Watch on YouTube