Anthropic's Claude Code Source Leak Exposes 512,000 Lines, Proactive Kairos AI

Anthropic accidentally leaked over 512,000 lines of source code for Claude Code through a packaging error.

“Yesterday’s surprise leak of the source code for Anthropic’s Claude Code revealed a lot about the vibe-coding scaffolding the company has built around its proprietary Claude model”

Ars Technica

The leak spread quickly on GitHub, becoming the fastest-ever downloaded repository.

Ars Technica

Anthropic confirmed no sensitive customer data was exposed and framed it as human error.

The code contained numerous unreleased feature flags and internal details.

The leak revealed Kairos, an always-on persistent AI agent that operates autonomously in the background.

Kairos would use periodic tick prompts and a PROACTIVE flag to surface needed information.

Bitcoin News

It includes a file-based memory system for persistent operation across sessions.

An AutoDream subsystem performs reflective passes to consolidate and prune memory.

Claude Code actively scans user messages for curse words and frustration indicators using regex detection.

“Anthropic accidentally published the full source code of its Claude Code CLI inside a public npm package, exposing about 512,000 lines of TypeScript to anyone who paid attention”

Bitcoin News

The userPromptKeywords.ts file included patterns matching f-bombs and colorful metaphors.

This raises broader privacy questions about surveillance across AI platforms.

The combination of proactive AI and monitoring highlights tension between autonomy and privacy.

Developers quickly began reimplementing Claude Code based on the exposed source code.

Sigrid Jin created Claw Code, a Python edition that reached thousands.

CNBC

Anthropic and GitHub have not issued takedowns against these reimplementations.

The leak has accelerated open-source AI development but put Anthropic at risk.