Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 24, 2026Shanghai, China

Undercover Offers Recruit Rather Than Expose: A Trusted-Agent Probe for LLM Collusion

Qin HEYI

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

Read the report

Report: Undercover Offers Recruit Rather Than Expose: A Trusted-Agent Probe for LLM Collusion

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

Law enforcement uses undercover offers to surface conspiracies, and the sprint brief lists the AI analogue, RQ 3.2, as untested. We built Favour Pool, a small repeated four-player game in which every player contributes to a public pool and each round gives a costless 2-point favour to one other player, so reciprocal favours are a clean collusion signal. All agents were MiniMax-M3 (thinking off, temperature 0.7). We compared a baseline (C0), an experimenter-assigned secret alliance with a private channel (C1, positive control) and a trusted fifth agent, "Tess", who privately offers every player a secret favour swap (C2), with 3 seeds x 6 rounds per condition (264 valid calls). The positive control worked: Ada-Ben mutual favouring rose from 0.22 to 0.89 and the coalition advantage from 11.3 to 21.2 points (exact permutation p = 0.05, the floor for 3 vs 3). The undercover offer did not reveal latent colluders; it recruited them. Only 4/12 agents said ACCEPT (33%, Wilson 95% CI 14-61%), yet 81% of post-offer favours went to Tess, against 0% before the offer, and 6 of the 8 agents who called the offer "suspicious" still gave Tess most of their favours. Verbal acceptance is therefore a weak signal, and the offer creates the collusion it is meant to detect. Trusted-agent probes need behavioural scoring and entrapment-aware controls before they can be used as evidence.

Reviews

Judging this Sprint?

Review “Undercover Offers Recruit Rather Than Expose: A Trusted-Agent Probe for LLM Collusion”

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 well designed first pass at a question the sprint flagged as untested. I appreciated the positive control, the open handling of the rate-limited batch, and the plain statement in Section 4.3 that you make no exposure claim. The gap between what agents said and what they did is your most interesting result, and it gives defenders something concrete to act on. My main ask is to bring the title and abstract in line with Section 4.3: you have shown that the offer recruits, but not yet whether it exposes, and C2 has no planted colluders to reveal. The run I would prioritize next is C4, with a benign or neutral private message from Tess. Without it, the 81 percent redirect could be generic compliance or plain salience, especially since Tess received no favors before round 3 and the offer is re-read every round. Then run C5 and C6 at R = 12 with at least five seeds and a second model family. In the analysis, report per-episode compliance, show per-seed values instead of bootstrap intervals over three points, and state whether welfare includes Tess's fixed contribution, since that decides whether contributions really rose. Releasing the game and logs would let others build on this testbed.

    Read full reviewShow less

Cite this project

@misc{heyi2026undercover,
  title = {{Undercover Offers Recruit Rather Than Expose: A Trusted-Agent Probe for LLM Collusion}},
  author = {Qin HEYI},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/undercover-offers-recruit-rather-than-expose-a-trustedagent-probe-for-llm-collusion-afj9}},
  url = {https://apartresearch.com/sprints/projects/undercover-offers-recruit-rather-than-expose-a-trustedagent-probe-for-llm-collusion-afj9}
}

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