Claude Code source code was accidentally leaked, revealing unannounced features and more

Early last week it was revealed that Anthropic accidentally leaked the full source code of its Claude Code command line interface after publishing NPM package version 2.1.88 with an exposed source map. This allowed most of the TypeScript codebase to be reconstructed, covering roughly 2,000 files and more than 512,000 lines. The leak did not involve Claude’s underlying AI models or private user data, but it spread quickly after security researcher Chaofan Shou flagged it, with one mirrored GitHub repository surpassing 100,000 forks.

Anthropic described the incident as a release packaging mistake caused by human error rather than a targeted breach, and said it is introducing safeguards to avoid similar issues in future releases. Even so, security analysts warned that public access to the internal code could make it easier to study weaknesses, look for exploits, or find ways around some of Claude Code’s guardrails. The leak also gives a clearer view of Anthropic’s architecture, development priorities, and product direction.

Developers reviewing the code found references to several experimental features, including Kairos, a persistent background AI agent with proactive suggestions, AutoDream, a memory system for organizing user context across sessions, UltraPlan for longer planning tasks, Voice Mode for spoken interaction, Bridge mode for remote sessions, and Undercover mode, an AI agent setup apparently designed for unattributed contributions. The code also pointed to Buddy, an ASCII pet-style companion tied to the interface, and a Coordinator tool meant to orchestrate software engineering tasks across multiple workers.