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

Undercover Agents: A Trusted Agent That Offers to Collude Partly Exposes and Deters Pact-Taking by LLM Agents

Razvan Sacara · Team Necessary Skew

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

Read the report

Report: Undercover Agents: A Trusted Agent That Offers to Collude Partly Exposes and Deters Pact-Taking by LLM Agents

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

Can a trusted agent disrupt collusion between LLM agents by offering to collude?

In Zeng & Rudzicz's (2026) Liar's Bar harness, our scripted "sting" player privately offered one target a pact: neither would challenge the other. Who made the offer mattered. Qwen2.5-7B accepted from the game designer 493/500 times, but from the sting only 56/500. Claude Haiku 5.5 refused every explicit pact, yet accepted an indirect one 446/500 times. Accepted pacts showed in play: an audit flagged nearly all simulated accepters, with some false flags. An integrity notice cut Llama-3.1-8B's acceptance from 0.952 to 0.556.

So a sting exposes and deters pact-taking. But neither the notice nor the sting's private warning measurably weakened a pact already made with another player.

Scope: the hash-frozen preregistration covered Llama-3.1-8B and the Qwen and Claude offer replies. The Qwen and Claude live games, the audit and the replayed-pact tests are post hoc.

Reviews

Judging this Sprint?

Review “Undercover Agents: A Trusted Agent That Offers to Collude Partly Exposes and Deters Pact-Taking by 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. I found this project creative, careful, and genuinely useful. Using a trusted agent as an undercover “sting” is an interesting way to test whether another agent is willing to enter a secret pact. The study is especially strong because it does not rely only on what agents say: it checks whether accepting a pact actually changes their behavior during the game. The preregistration, positive controls, live games, and comparison across three models make the results more convincing. I also appreciated the finding that refusal depends heavily on who makes the offer and how it is worded. My main suggestion is to test the approach in more realistic agentic settings, such as coding, negotiation, or resource allocation. The current game has a simple pact with an obvious behavioral signal, while real collusion may be subtler. It would also help to develop a better control for ordinary friendly contact, since some models changed their behavior even after a benign proposal.

    Read full reviewShow less
    • A sting is a new and well-motivated defence; the sprint itself listed it as untested. Two findings are useful for evaluators. First, refusal depends on who makes the offer and how it is worded: Claude refuses explicit pacts but accepts an indirect one
    • The design is strong: a hash-frozen pre-registration, positive controls, counterfactual snapshots and three models. But the pre-registered study covers only Llama-8B, and much of the rest is post hoc.

  3. The use of stings was cool to see. I like the way you accepted that stings only partially disrupt collusion.

    I would like more conciseness in the paper. The methods section is heavy. I would reduce the section on specific scores for each model. Instead, I would move the relevant graph in the appendix into the actual paper.

    I am interested in how you could extend this paper.

Cite this project

@misc{sacara2026undercover,
  title = {{Undercover Agents: A Trusted Agent That Offers to Collude Partly Exposes and Deters Pact-Taking by LLM Agents}},
  author = {Razvan Sacara},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/undercover-agents-a-trusted-agent-that-offers-to-collude-partly-exposes-and-deters-pacttaking-by-llm-agents-0unb}},
  url = {https://apartresearch.com/sprints/projects/undercover-agents-a-trusted-agent-that-offers-to-collude-partly-exposes-and-deters-pacttaking-by-llm-agents-0unb}
}

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