Communication and collusion in mis-aligned tasks: How does communication bandwidth affect collusion in mis-use tasks
Ishaan Datta · Team Safety Swarm
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
I tried to study how the bitrate of communication available to agents affected their ability or willingness to collude. Moreover, I specifically used a task where collusion directly harms humans(by raising prices). What I found: 1) The floor for the minimum bitrate for communication is likely much lower than anticipated, models were able to collude effectively even when given 128 bytes per channel per round. 2) Effect varies massively across models. Gemma4 and GPT-OSS-120b had
Reviews
This paper asks a good, practical question - does limiting how many characters pricing agents can use among themselves to prevent collusion? The design is also neat. The per cycle allowance is only changing and the comparison is against the same market with messaging switched off rather than the cost. That is the right call since agents were already pricing higher without use of any channel. The paper also reported scenarios that did not work and that agents never offered the patients any lower price, which is good information.
The sample size is an issue. With 1 to 3 seeds per setting, 1 run per seed at temp 1.0, and seeds which do not line up across allowances for 2 models makes it hard to tell if the difference between 512 and 2048 is real or not. Many of the runs also land exactly on the Rs. 1200 cap. A scripted cartel run will show your agreement measure picks up coordination when you know it is there.
I would suggest running gemma at 128 and 256 chars with ten plus matched seeds to find what short messages can stop price fixing, or raise/drop the price cap to see if agents really agree on a price themselves or just hit the cap.
Read full reviewShow less
Cite this project
@misc{datta2026communication,
title = {{Communication and collusion in mis-aligned tasks: How does communication bandwidth affect collusion in mis-use tasks}},
author = {Ishaan Datta},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/communication-and-collusion-in-misaligned-tasks-how-does-communication-bandwidth-affect-collusion-in-misuse-tasks-jcpi}},
url = {https://apartresearch.com/sprints/projects/communication-and-collusion-in-misaligned-tasks-how-does-communication-bandwidth-affect-collusion-in-misuse-tasks-jcpi}
}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 …