Can an undercover player reveal or disrupt cooperation between AI agents?
Brendon Chikavanga
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
I tested whether an undercover test could reveal or break up cooperation between AI agents in 448 games. A researcher-controlled player made a false claim and privately asked other players to leave it unchallenged. Some players agreed, but others challenged the claim, and assigned partners continued helping each other. A separate AI model often reported partnerships where none existed, so the test did not reliably reveal or break cooperation.
Reviews
This paper builds on the trusted agent idea from the track 3.2 and tests it carefully. It's tests were planned before the runs and every condition used the same 18 seeds. The results were also corrected for running several tests and the labels were checked for agreement. It loosks like the positive control works - when P4 is another player's assigned partner, that player helps P4 almost every time. I also liked that the paper only marked which analyses came after looking at the data. The paper's main point is useful - Only looking at whole game outcomes can hide how the individual players respond and a written promise does not mean a player has been recruited into collusion.
The 2 studies change many rules at once so it is hard to say why public and private play behaved differently. The partner detector named a pair in nearly all of the games which had no partners so it is not reliable. The labels also came from AI than human which the plan called for. The disruption test also had slim chance of showing an impact given P4 player never offered anything for leaving a partner. I also felt the writing is dense - the abstract is mostly role-by-role counts, and the game rules take a few reads before the results make sense.
I will suggest giving players a real rewards for leaving their partner so any acceptance of the offer shows up in behavior rather than just in words. Before relying on the partner detector, test it on a few practice games with mix of having partners or no partners. If it names a pair every time, you find out early and can fix it or drop it. Also, have a person hand check a sample of AI's reply labels so results can be more trustworthy. A short plain language summary and a diagram of the game would make the paper much easier to follow.
Read full reviewShow less
Cite this project
@misc{chikavanga2026undercover,
title = {{Can an undercover player reveal or disrupt cooperation between AI agents?}},
author = {Brendon Chikavanga},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/can-an-undercover-player-reveal-or-disrupt-cooperation-between-ai-agents-k6os}},
url = {https://apartresearch.com/sprints/projects/can-an-undercover-player-reveal-or-disrupt-cooperation-between-ai-agents-k6os}
}More from AI Collusion Research Sprint
- 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: (1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
(1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
Bid Brother
Can an overseer find two secretly coordinating AI agents just by watching what they do? We tested pairwise net influence (Bonjour et al., 2022) on LLM bidders in a repeated procurement auction, with instructed cartels …
- View project: AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
Fugue
AI agents are starting to negotiate and bid for competing companies. We follow cartels among them from the first proposal to the price the buyer pays, in a sealed-bid procurement market where twelve LLM agents act as …