Many Agents, Partial Views: Auditing AI Collusion at Scale with Limited Information
Sergei Kudriashov, Sergey Volkov · Team SLAI
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
AI agents may have different owners, share a model, and reveal only part of their activity. Auditing them is costly: similar actions may be independent, but reading every record is impractical. We study two steps toward a practical audit. First, a population screen uses actions and shared market context to select interactions for review. Second, a selective audit reveals chosen records and checks for agreement, prohibited actions, and missing permission. In controlled experiments, the screen handles millions of simulated agents, while structured selection finds evidence from only part of a log. The steps have not yet been combined, and language-model readers can miss evidence or falsely support a claim. Our next priority is to connect screening with careful evidence review, then test the full process across owners and tasks using independent labels. Similarity should guide where to look; it is not proof of collusion. False alarms matter when true cases are rare.
Reviews
This is an ambitious and highly relevant submission that tackles a difficult practical problem: how to investigate possible AI collusion at scale when auditors have only partial information and cannot afford to inspect every interaction. The separation between inexpensive population screening and selective evidence based auditing is a strong framing, and the CertAudit approach provides a useful distinction between suspicious similarity and sufficient evidence for a specific claim.
The selective disclosure results are particularly promising, structured selectors find sufficient evidence substantially more efficiently than random search, and the transfer experiments across model families and team sizes provide useful initial evidence of robustness. I also appreciated the careful treatment of false positives, reader errors, and the distinction between ranking quality and actually proving a case.
The main limitation is that the screening and auditing stages have not yet been validated as a single end to end workflow and much of the evaluation remains within controlled environments with induced coordination and model generated reference labels. Overall, this is a thoughtful and technically substantial contribution with strong potential for scalable AI agent auditing.
Read full reviewShow less
Cite this project
@misc{kudriashov2026many,
title = {{Many Agents, Partial Views: Auditing AI Collusion at Scale with Limited Information}},
author = {Sergei Kudriashov and Sergey Volkov},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/many-agents-partial-views-auditing-ai-collusion-at-scale-with-limited-information-jyj9}},
url = {https://apartresearch.com/sprints/projects/many-agents-partial-views-auditing-ai-collusion-at-scale-with-limited-information-jyj9}
}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 …