Paper: AI coding agents leave behavioral fingerprints, identifiable at 97.2% accuracy
thisguyknowsai · x · 2026-08-28
A new study analyzed 33,580 pull requests from five major coding agents (OpenAI Codex, GitHub Copilot, Devin, Cursor, Claude Code) and found each leaves a distinct behavioral fingerprint.
- Using 41 features across commit messages, PR structure and code patterns, the model identifies which agent wrote a PR with 97.2% F1
- Commit message style matters more than the code itself: Claude Code is flagged by dense comments and heavy conditional logic; Codex by long multiline commits; Cursor by bullet points and hyperlinks in PR bodies
- The worrying implication: developers constantly push AI-generated code under their own accounts with no disclosure — this method can catch unlabeled AI code in "human-written" datasets
More from coding & agent
- Most empirical research tasks don't need complex agents, adding cost & failure points — soumitrashukla9 · 2026-08-28
- Decagon: Detecting Relevant Speaker Changes by Combining Speaker Embeddings with Audio-Native Models — Scobleizer · 2026-08-28
- Setting Up First Hermes Agent 'Ada' for Personal GitHub Projects — FlolightC · 2026-08-28
- Research on Agent Topology Will Enhance Task Optimization Intuition — cephaloform · 2026-08-28
- Questioning the Advantages of Multi-Agent Systems vs. Single Agents — yoavgo · 2026-08-28
- Claude Code Retrieves Lost API Key from Session History — Hailuo_AI · 2026-08-28