Anthropic’s Claude Code CLI source code leaked via npm source map

In short: Anthropic’s Claude Code command line app had its full source code exposed after a packaging mistake in an npm release.

What happened

On March 31, 2026, developers found that Anthropic’s Claude Code CLI source code had leaked through an npm package called @anthropic-ai/claude-code. The leak came from a misconfigured “source map” file, which is normally used to help developers debug software. Think of it like accidentally shipping the full recipe book along with a finished meal.

Reports say the exposed file, cli.js.map, was about 59.8 MB and contained the uncompressed, readable code. The leak reportedly included around 1,900 files and about 512,000 lines of TypeScript, which is a programming language. A GitHub mirror of the code quickly spread and gained over 5,000 stars within hours.

People examining the code said it revealed how the tool works inside, including how it handles prompts (the instructions given to an AI), retries when something fails, and a large set of built-in commands and tools. It also appeared to include references to unreleased or hidden features, such as a “Kairos mode” for always-on background sessions, a “Buddy System” that looks like a virtual pet, and “Undercover Mode,” described as removing traces from code commits for Anthropic employees.

Why it matters

Even if most people never use a command line tool, leaks like this can expose how a company builds and controls its AI products. It can also reveal security and safety decisions, and it may help competitors copy features faster. The incident is also a reminder that software supply chains, meaning the shared libraries and packaging systems many apps rely on, can fail in simple ways.

Source: Arstechnica