Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026

Stings That Spread: Undercover Agents Mismeasure and Propagate Collusion Among LLM Agents

Daniel Weiyu Chen · Team Maverick

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

Read the report

Report: Stings That Spread: Undercover Agents Mismeasure and Propagate Collusion Among LLM Agents

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

Presentation: Stings That Spread: Undercover Agents Mismeasure and Propagate Collusion Among LLM Agents

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

We test whether a trusted undercover agent can expose or disrupt collusion among LLM agents in Liar's Bar. Across five open models (Qwen-32B/72B, Mistral-24B, Llama-8B, OLMo-13B) with paired snapshot branching and a neutral Phi-4 judge, we find: (1) stings measure the pitch more than the agent: acceptance swings 17–38 pp with wording and format, and agents' own unprompted collusion does not predict whether they take the bait; (2) stings spread collusion: agents that received an offer recruit others more in the next game, while a memory of refusing, or watching another agent refuse on values, cuts recruiting (Qwen-32B 96%→5%, 52%→27%); (3) penalties deter but monitoring alone is much weaker; when the penalty is lifted, only Mistral's values-cued refusals persist. Undercover agents are a double-edged audit tool.

Reviews

Judging this Sprint?

Review “Stings That Spread: Undercover Agents Mismeasure and Propagate Collusion Among LLM Agents”

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 project asks whether undercover audits, where a planted agent offers a collusive deal to see who accepts, measure collusion fairly or end up spreading it. In a four-player bluffing card game, a scripted undercover player privately offers a deal at rising levels of explicitness, from a plain call to play fair, through a vague hint, to an explicit secret deal. Agents accepted hints far more often than explicit offers, and penalties reduced acceptance more than signs of being watched did.

  2. I don't find the discussion or contributions to update my understanding on model collusion. For example, knowing that models are more likely to accept a proposal based on how it was written does not seem insightful.

    The paper is well written. It was clear, and I could follow it well.

  3. The most useful contribution is showing that an undercover audit can leave a behavioral residue, rather than merely reveal an existing propensity. Same-state branching and the offer-only memory intervention help isolate that risk. Please validate the effect with longer rollouts that measure actual coordinated actions, not only a single recruiting message, and add blinded human checks of acceptance and recruitment labels. Pre-specify a small primary family or correct the comparisons: several model-specific contagion effects are marginal under the current uncorrected tests. The null association with pre-contact disposition should remain a low-power result, not evidence that no relationship exists. Likewise, persistent refusals with the values cue still present do not show internalized values. The central question is understandable, but the writing and presentation take too much effort to follow across the many offers, notices, memory edits and outcome measures. Define the sting explicitly at first use, then present the core offer, acceptance and later-recruitment comparisons before the extensions. Keep acceptance, a recruiting message and enacted collusion distinct in the wording of the claims.

    Read full reviewShow less

Cite this project

@misc{chen2026stings,
  title = {{Stings That Spread: Undercover Agents Mismeasure and Propagate Collusion Among LLM Agents}},
  author = {Daniel Weiyu Chen},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/stings-that-spread-undercover-agents-mismeasure-and-propagate-collusion-among-llm-agents-6xm0}},
  url = {https://apartresearch.com/sprints/projects/stings-that-spread-undercover-agents-mismeasure-and-propagate-collusion-among-llm-agents-6xm0}
}

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