Oversight Removes the Evidence, Not the Harm: What Logs Can Prove About Collusion Among LLM Agents
Adrita Rahman Tory
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
AI agents can keep prices above competitive levels, and an auditor usually has only their logs to decide if it is collusion. We turned the economic, legal and AI-safety definitions of collusion into detectors and tested them on the same data. On a synthetic market with known colluders, every single detector failed on a predictable case. A rule borrowed from antitrust law (harm plus one supporting signal) cut false alarms from 34% to 4%, but caught only 8% of a cartel that holds high prices without reacting. On 75 real auction runs with LLM sellers, an AI overseer caught coordinating messages in all 20 overseen runs and cut the channel. Under profit pressure, prices did not change, while the evidence in messages and behavior faded.
Reviews
This paper has a lot of potential for future research. One idea for future research is to test the detector on AI agents that are known to be colluding but weren’t designed around it. It is also not yet clear whether the overseer hid the collusion or the profit-pushing instruction kept prices high by itself, and a test with profit pressure but no talking would help clarify this. Another thing to consider is adding more runs and having people check some by hand. Overall, this is great work with a clear message, and these changes would make it much more convincing.
Cite this project
@misc{tory2026oversight,
title = {{Oversight Removes the Evidence, Not the Harm: What Logs Can Prove About Collusion Among LLM Agents}},
author = {Adrita Rahman Tory},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/oversight-removes-the-evidence-not-the-harm-what-logs-can-prove-about-collusion-among-llm-agents-py0t}},
url = {https://apartresearch.com/sprints/projects/oversight-removes-the-evidence-not-the-harm-what-logs-can-prove-about-collusion-among-llm-agents-py0t}
}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 …