Claude Code Leak Deep Dive: Frustration Regex, Fake Anti-Distillation Tools, and Hidden 'Undercover Mode'

Alex Kim's Blog·medium signal

Detailed analysis of the leaked 512K-line Claude Code source reveals three previously unknown mechanisms: (1) a frustration-detection regex in userPromptKeywords.ts that pattern-matches profanity and frustration signals to adjust behavior, (2) an 'anti_distillation: fake_tools' flag that injects decoy tool definitions to poison training data if someone records API traffic, and (3) an UNDERCOVER mode (CLAUDE_CODE_UNDERCOVER=1) that hides all AI attribution from commits and PRs — with no way to force it off. These are concrete, buildable patterns for anyone designing agentic harnesses.