Claude Code source code LEAKED

TL;DR

An Anthropic engineer accidentally leaked Claude Code's source map, exposing 600,000 lines of TypeScript and unreleased features before it spread to 44,000 GitHub forks.

Key Points

• 1.The leak originated from an accidentally included 60MB source map file. It could be converted into ~200,000 TypeScript files and 600,000 lines of readable code; one GitHub mirror was forked 42,000 times within hours before Anthropic pulled it.
• 2.The leak exposed a full product roadmap of unshipped features hidden behind false flags. These include Chyros (a background GitHub-monitoring agent), AutoDream (a memory-consolidation agent), Ultra Plan (a 30-minute remote deep-planning session), a coordinator multi-agent swarm mode, persistent cross-session memory, real browser control, and a real-time voice mode.
• 3.A user rapidly converted the leaked TypeScript codebase to Python using OpenAI's Codex within hours. This raises a major legal gray area: AI-assisted 'clean room engineering' may effectively erase copyright by replicating functionality in entirely different code.
• 4.A full virtual pet (Tamagotchi-style) system was hidden in the code, triggered by the /buddy command. It features 18 species, a gacha rarity system with 1% legendary drop rate, shiny variants, hats, and three stats: debugging, chaos, and snark.
• 5.Claude actively monitors user language to detect frustration during sessions. There is also a mysterious 'undercover mode' flagged by an AI researcher that is still being investigated, and 187 different loading spinner verb variations.
• 6.The Mythos model codenamed Capibara is confirmed in the source code, and Anthropic appears to be building agentic crypto payment infrastructure. No model weights, training secrets, customer data, or API credentials were part of the leak.