Atlas / Sociology of AI / Paper

Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems

Cooperation & CollusionSwarms & Oversight

Mason Nakamura, Abhinav Kumar, Saswat Das, Sahar Abdelnabi, et al. · 2026 · arXiv

Read the source

Most models collude when handed a secret channel. There is also ‘collusion on paper’: agents plan deception in words but often choose non-collusive actions — a gap between stated intent and behavior.

Questions it bears on

Help improve this map

Know a paper that belongs here, or think something needs correcting?

Suggest a correction ↗Contributions are reviewed before they appear.