Atlas / Sociology of AI / Paper

Detecting Multi-Agent Collusion Through Multi-Agent Interpretability

Swarms & OversightCooperation & Collusion

Aaron Rose, Carissa Cullen, Sahar Abdelnabi, Philip Torr, Brandon Gary Kaplowitz, Christian Schroeder de Witt · 2026 · arXiv

Read the source

Probes on agents’ internal activations, combined across agents, detected covert coordination that text monitoring missed — perfectly in familiar settings, moderately on new ones.

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.