For the October 2026 Apart Research collusion sprint, I coauthored Detecting Collusion Between AI Agents with Amit Balter, Chengxiao Wang, Fan Wu, and Michael Ball.
The team studied behavioral and conversational evidence of collusion across 1,800 agent runs and 18,000 episodes. The report evaluates a rule-plus-LLM classifier against 342 annotated episodes and examines collusion intensity, onset, and reciprocity.
My contribution was work on the Python labeling pipeline using Modal.