AI Startup Leaks Code Again, Revealing Hidden Features

Anthropic, the AI company behind Claude, has inadvertently released the full source code for its Claude Code CLI tool to the public npm registry. This incident, occurring again after a similar event in early 2025, involved a 57 MB source map file that exposed over 2,300 TypeScript files.

The leaked codebase offers a detailed look at the client-side interface, agent runtime, and tooling systems. Developers examining the files uncovered references to an unannounced 'Capybara' model family and experimental features. Notably, a hidden command suggests the existence of a Tamagotchi-style AI companion system with unique traits.

The code also details Anthropic's telemetry practices, showing that Claude Code logs user frustration signals, such as swearing or repeated commands. While safeguards prevent the transmission of user code, this comprehensive data collection is routed through Datadog. Users have the option to disable telemetry.

This exposure provides unprecedented transparency into Claude Code's permissions model, including logic for tool approvals, file access, and execution boundaries. The complete system prompt is also visible, offering insights into how Anthropic guides the AI's behavior and enforces its limits.

Developers have already begun forking the exposed code, with some working on stripped-down or alternative versions. Anthropic has not yet issued a public statement but is expected to address the security lapse by rotating keys and refining its release process. The incident highlights the challenges in maintaining code security, especially with tools used by developers.