Anthropic Accidentally Exposes Claude Code's Secret Code on npm

Anthropic, the company behind the AI model Claude, has recently faced an unexpected leak of its source code. This leak occurred due to a publishing error on the npm platform, where a .map file revealed the complete code of Claude Code, the command-line tool for developers. Although the model weights were not compromised, this blunder exposes the company's strategic plans.

On March 31, users on the X platform reported this accidental leak. Version 2.1.88 of the package "@anthropic-ai/claude-code" contained a .map file, which normally serves as a blueprint and should remain private. By inadvertently uploading it, Anthropic revealed the full plans for its interface. What was leaked is not the brain of the AI (the model weights), but the source code of a command-line tool written in TypeScript.

An Ambitious Roadmap for 2026

The leaked source code has unveiled several projects that Anthropic plans to develop by 2026. Among these projects is ULTRAPLAN, a system designed to delegate complex tasks to a remote instance of Claude. Another project, KAIROS, aims to transform Claude into a persistent agent with long-term memory, capable of consolidating its logs through a "night dreaming" process.

Additionally, the Coordinator mode will allow the tool to clone itself to have multiple sub-agents working in parallel, under the supervision of a "Master Claude." Anthropic is also testing a new language model named Capybara and developing frustration telemetry to measure user annoyance levels.

The BUDDY Project: An Interactive Virtual Companion

Among the revelations, the BUDDY project stands out for its originality. Inspired by the famous Tamagotchi, this virtual companion will come to life in the command interface. Initially planned as an April Fool's joke, BUDDY could become a permanent feature as early as May. This randomly generated companion will be able to take on 18 different forms, such as a duck, a capybara, or a robot, and will react to developers' actions.

The BUDDY project is teased from April 1 to April 7, 2026, and will feature procedurally generated statistics. This playful touch could attract internet users and further enhance their loyalty.

A Delicate Time for Anthropic

This leak comes at a time when Anthropic is already at odds with the Pentagon regarding the use of AI in the military domain. Furthermore, at the end of March, another leak revealed Claude Mythos, the company's upcoming AI model. While this error does not compromise the core of the AI, it nonetheless exposes valuable industrial secrets.