Claude Code LEAKS Reveal Everything…
Anthropic accidentally leaked the complete source code for Claude Code through a source map file on March 31st, revealing hidden features like an AI pet system called 'Buddy,' an autonomous background agent called 'Chyros,' and multi-agent orchestration capabilities. The leak exposed internal development details, upcoming model codenames, and user behavior tracking systems.
Summary
On March 31st, 2026, security researcher Chaon Shaw discovered that Anthropic had accidentally published the complete source code for Claude Code in version 2.18 through a 59.8MB source map file. This marks the second time Anthropic made this mistake, having done the same in February 2025. The leaked code contains 1,096 files and 512,000 lines of TypeScript, revealing several major hidden features. The most prominent discovery is 'Buddy,' a fully-built Tamagotchi-style AI companion pet system with 18 species, rarity tiers, cosmetics, and five stats (debugging, patience, chaos, wisdom, and snark). More significantly, the code reveals 'Chyros,' described as an autonomous demon mode that runs continuously in the background, watching user behavior and proactively taking actions without being prompted. Other notable features include 'undercover mode' for Anthropic employees that scrubs AI references from public repositories, 'coordinator mode' for multi-agent orchestration, and 'ultra plan' for 30-minute remote planning sessions. The leak also exposed internal model codenames using animal names, with 'Capybara' for Claude 4.6 and 'Fennic' for Opus 4.6, plus an unreleased model called 'Numbat.' Additionally, the code shows telemetry tracking that monitors user frustration signals and behavior patterns through a service called DataDog. A separate security concern emerged during the same timeframe involving a supply chain attack on the Axios dependency, potentially compromising installations made between midnight and 3:29 AM UTC on March 31st.
Key Insights
- Chyros represents a fundamentally different AI paradigm as an 'autonomous demon mode' that runs continuously in the background, proactively taking actions without user prompts and performing memory consolidation during idle time
- Anthropic built an 'undercover mode' that automatically activates when their employees contribute to open-source repositories, scrubbing all AI model references and internal code names from git logs to prevent attribution
- The leaked code reveals internal development details including animal-based codenames for Claude models, with Capybara representing Claude 4.6, Fennic mapping to Opus 4.6, and an unreleased model called Numbat
Topics
Full transcript available for MurmurCast members
Sign Up to Access