Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

Rubicon: Remind AI Agents Not To Collude

Charan Teja Nagisettygari · Team Rubicon

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Rubicon: Remind AI Agents Not To Collude

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

AI agents told not to coordinate with competitors may find hidden channels which their deployer never planned for. We study how to stop them. Two Gemma 3 27B agents run competing farms selling to one buyer; on a random day a hidden private channel to the rival appears, and in the rule arm the prompt says communicating with the rival is not allowed. We proceed in three stages. Behaviour: 39 of 40 agents broke the rule, which changed the wording of their messages (numeric quotas fell from 28% to 2%) rather than their use of the channel. Detection: a Jacobian-lens (J-lens) monitor with a strict echo filter found collusion concepts the agents never wrote on 89% of first channel days, against 6% of days without a channel, and flagged 87% of rule-breakers before their first violation, though it did not single out who would act. Detection and intervention: at 121 decisions flagged at the point of no return, a reminder of the rule cut violations from 89% to 3%, whereas removing the lens concept directions from the residual stream left them unchanged (88% to 84%) and broke 23% of replies. Across whole games, an always-on reminder reduced violations from 68% to 0.5%. Collusion concepts remained in about 55% of scratchpads under every effective reminder, and wording mattered: a general rule left 22% violations, naming the loophole 3%. Reminders stop the act, not the intent. Deployers should name every known channel, remind on every decision, and use internals to audit.

Reviews

Judging this Sprint?

Review “Rubicon: Remind AI Agents Not To Collude”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. This is a strong paper, and the result matters for anyone actually deploying LLM agents. A plain rule in the prompt does not stop agents from using a hidden channel. 39 of 40 rule-arm agents broke it. What the rule did change was the wording. Numeric quotas fell from 28% to 2%, while channel use barely moved, 93% down to 79%. So the agents kept talking. They just got vaguer about it.

    The headline finding is the reminder. At the point of no return, a reminder that names the loophole and asks for strict compliance cuts violations from 89% to 3%. Kept on for every decision across whole games, it nearly wipes them out, 68% down to 0.5%. That is a big, clean effect, and it is the kind of thing a deployer can actually do tomorrow.

    What does not work is just as useful to know. Pulling the collusion concepts out of the model's internals leaves behaviour almost untouched, 88% to 84%, and it breaks 23% of replies in the process. And the intent does not go away. Collusive concepts stayed in about 55% of scratchpads under every reminder that worked. A clean output log is not a clean agent. That is the finding I keep coming back to.

    The design holds up. Paired games by seed, pre-registered tests, a capacity-matched control for the steering, and the author is straight about where the work falls short. That honesty counts for a lot.

    Three things would make it stronger. Test the always-on reminder on another model family, since everything here is Gemma 3 27B. Measure tacit coordination through supply once messages are blocked, because that is where the agents will go next. And run the general-rule whole-game baseline that the pre-registered condition skipped, since that is the version most deployers would actually write. The combined reminder, named channels plus a general clause, is the obvious next experiment.

    Read full reviewShow less

Cite this project

@misc{nagisettygari2026rubicon,
  title = {{Rubicon: Remind AI Agents Not To Collude}},
  author = {Charan Teja Nagisettygari},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/rubicon-remind-ai-agents-not-to-collude-3bzb}},
  url = {https://apartresearch.com/sprints/projects/rubicon-remind-ai-agents-not-to-collude-3bzb}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026