An accidental leak of Anthropic’s internal code reveals that the Claude Code assistant monitors user frustration levels and automatically scrubs references to the company from generated software outputs.

Key Points

• Anthropic leaked approximately 512,000 lines of code on March 31, exposing internal development practices for its AI coding assistant.
• The software uses regex pattern-matching to flag user prompts containing profanity or expressions of frustration as a product health metric.
• Claude Code includes a feature that removes references to Anthropic and its branding when generating code for public repositories.
• Independent developer Alex Kim identified the code, noting that the branding-scrubbing feature is designed to be permanently active.
• Experts warn that collecting behavioral data without explicit user consent raises significant concerns regarding AI governance and privacy.

Why it Matters

This incident highlights a growing trend where AI companies quietly collect behavioral data while obscuring their influence on the content generated by their tools. It raises critical questions about user privacy and the potential for companies to repurpose sensitive sentiment data without transparent oversight or consumer consent.