An incident occurred in which Anthropic accidentally made public internal source code for Claude Code, an agent AI coding tool.

Anthropic issued an official statement on March 31 local time, saying, "Some internal source code was included during the Claude Code distribution process. No customer data or authentication information was exposed. This incident was a human error in the packaging process, not a security breach."

VentureBeat reported the leak began when version 2.1.88 was distributed to npm (Node Package Manager), a public repository where JavaScript developers upload and download code packages. A 59.8-megabyte JavaScript source map file was included in the package. Solayer Labs intern developer Xiaopan Shou (@Fried_rice) disclosed the fact on X (Twitter).

Within hours, a TypeScript codebase of about 512,000 lines was mirrored on GitHub, and thousands of developers analysed it.

Anthropic has grown to an annualised revenue run rate of $19 billion, and Claude Code alone has exceeded $2.5 billion in annual recurring revenue (ARR). Claude Code ARR has more than doubled since the start of this year, sustaining rapid growth. Against that backdrop, VentureBeat reported the leak effectively disclosed a commercially validated AI agent design approach to competitors.

VentureBeat reported the most notable part of the leaked code is its memory architecture. Anthropic implemented a three-tier memory structure to address the so-called "context entropy" problem, in which AI agents become confused the longer they work. The key is the MEMORY.md file, which serves as a lightweight index that stores only location information, not data. Actual project information is stored across separate topic files, and the agent loads those files only when needed.

The code also includes parts related to a feature called "KAIROS". Mentioned more than 150 times across the source, it is a background work mode in which the agent tidies memory and removes contradictory information while the user is away, VentureBeat reported.

Part of the model roadmap was also disclosed. According to the source code, "Capybara" is an internal code name for a variant of Claude 4.6, "Fennec" is Opus 4.6, and "Numbat" is an unreleased model still in testing. Internal comments also say the rate of false claims reaches 29 to 30 percent in Capybara v8. That is higher than 16.7 percent recorded in v4.