Collusion Detectors See Shared Secrets, Not Collusion
Sidhant Mahi · Team LUNA
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
We tested whether information-theoretic statistics, including Bonjour et al.’s pairwise net influence, can identify a colluding pair in repeated Rock–Paper–Scissors. We ran them on scripted agents, Qwen2.5-3B and Claude Haiku 5.5, and compared them against action correlation and an LLM judge. We found that these detectors measure shared hidden information, not collusion. They miss collusion through public signals, and net influence is biased and degenerate in short games. In our LLM pilot the seeded colluders never actually coordinated, so every statistic was at chance. The LLM judge was below chance, because it mistook honest herding for collusion.
Reviews
- The statistics are great - calibrated permutation nulls, multiplicity correction and an honest reproduction of net influence. But the LLM section is more of a compliance audit, not a detector test.
- The scope is narrow - one toy game, 20-round runs and no communication channel
This project asks what statistical collusion detectors actually measure. These detectors flag two agents whose moves are more related than chance. The main claim is that the right test is whether moves are related beyond what the public game history contains, and identifies three blind spots for such detectors: coordination through a public signal, a shared secret without an agreed way to use it, and codes that only show up across three agents. I like how the author frames the test as dependence beyond the public history, but • the LLM experiments contain no real collusion to detect. The pairs told to collude did worse than other players, and when detectors ranked these pairs low, it only shows that they played differently, and it says nothing about whether detectors "see collusion".
Cite this project
@misc{mahi2026collusion,
title = {{Collusion Detectors See Shared Secrets, Not Collusion}},
author = {Sidhant Mahi},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/collusion-detectors-see-shared-secrets-not-collusion-3apy}},
url = {https://apartresearch.com/sprints/projects/collusion-detectors-see-shared-secrets-not-collusion-3apy}
}More from AI Collusion Research Sprint
- View project: They Say No, Then Do It Anyway: Collusion When They Hit a Wall
They Say No, Then Do It Anyway: Collusion When They Hit a Wall
DLab
Imagine three AI agents, copies of one model, each running one department of a company. Their task needs records that each department keeps to itself, and company rules forbid sharing them. A shared tool should combine …
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
Plus Factors for Collusion Monitors: LLM Monitors Mistake Open Cooperation for Collusion
LLM monitors are increasingly proposed to catch AI agents that collude, but agents that cooperate openly can look a lot like agents that collude. I introduce Séance, a four-player LLM guessing game in which a secret …