The Claude Code Source Leak Analysis

Anthropic leaked the full source code for Claude Code on April 1 (not a joke : ), 2026. This leak happened through an NPM package update. A large source map file revealed over 500,000 lines of code. This gives us a rare look at how advanced AI agents work.

The leaked code shows the inner workings of the tool. It reveals several secret features and internal tools that were not public.

Top Revelations from the Leak

Undercover Mode: Claude is instructed to hide its AI identity. It avoids mentioning Anthropic in code commits to look like a human developer.

Anti-Distillation: The system uses fake logic to stop other companies from training their models on Claude's data.

Frustration Detection: The code scans for swear words. This helps the system track when a user is unhappy with the service.

Hidden Models: The leak mentions new models. These include Opus 4.7, Capybara, and a model called Mythos.

Kairos Agent: This is a background tool. It keeps memories between sessions and performs tasks on a schedule.

Why the Leak Occurred

A bug in Bun.js likely caused the leak. This technical glitch allowed private files to be served in a public update (sourceMaps). Developers are already using this code to build open versions of the technology.

If you need help with software feel free to reach out to me via https://www.fortesglobalweb.nl. I am a Freelance Software developer. I automate workflows for businesses as a Senior Developer and founder of my AI platform, Word-flux (https://www.word-flux.com).