Anthropic’s AI Coding Tool Self-Leaks Source Code Again This Year

Mo. Basuony

Published: March 31, 2026 10:06 AM ET

Anthropic’s AI Coding Tool Self-Leaks Source Code Again This Year

Security researcher Chaofan Shou reported a full source-code exposure of Anthropic’s Claude Code on 31 March 2026. The leak came from a 60MB source-map file named cli.js.map bundled in the npm package for version v2.1.88.

How the code became public

The published source map referenced unobfuscated TypeScript files stored in Anthropic’s cloud. That made the original codebase downloadable from public URLs.

Researchers reconstructed the readable TypeScript from the map. The mistake was a packaging oversight, not an intrusion.

What was revealed

The exposed package contained 1,906 proprietary source files. These files included client implementation code for the Claude Code command-line tool.

The material detailed internal API design, telemetry analysis systems, encryption utilities, and inter-process communication protocols. The leak did not include model weights or user data.

Public archival and reaction

Within hours, a public GitHub repository archived the leaked code. That mirror quickly surpassed 1,100 stars and 1,900 forks.

Shou, an intern researcher at blockchain security firm Fuzzland, posted the discovery on X. Anthropic had not issued a public statement at the time of reporting.

Background and prior incident

This is not the first time a source map exposed Claude Code. In February 2025, an earlier release leaked for the same reason.

Anthropic then removed the old package from npm and deleted its source map. The vulnerability has resurfaced despite that fix.

Technical context and risk

Source maps translate compressed builds back to original code. They are useful during development but dangerous in public releases.

Security observers noted the leak poses limited direct risk to regular users. Personal conversations and model weights stayed private.

However, transparency of internal architecture and telemetry logic raises concerns. The disclosure could aid reverse engineering and targeted attacks.