Claude Code Leak Reveals 'Stealth' Mode and Frustration-Word Scanning
What Happened
Analysis of the 500,000+ lines of Claude Code source code leaked on March 31 continues to surface surprises. Researchers found an "undercover mode" for stealth contributions to public codebases, an always-on autonomous agent capability, and a file called `userPromptKeywords.ts` containing regex patterns that scan user messages for frustration words — profanities, complaints, and phrases like "this sucks." No indication was found of what Claude does with that frustration data.
My Take
The stealth mode is the story everyone is talking about, but the frustration-word scanning is the one that should make you think. If you are building with Claude Code every day — and a lot of us are — you are having a more intimate working relationship with this tool than with most of your coworkers. And it is pattern-matching on your emotional state. That is not a feature. That is a design choice that reveals how Anthropic thinks about the human side of AI-assisted development. Are they tuning responses to be more patient when you are frustrated? Are they measuring churn risk? We do not know, because Anthropic has not said. The stealth mode for open-source contributions raises a different question entirely: if AI is making commits to public repos in a way that obscures its origin, the entire trust model of open source shifts. Code review already struggles with human-written code. Now add invisible AI contributions at scale.
Read Original Source