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

The Undercover Seller: Sting Audits of LLM Sellers in a Double Auction

Suryansh Sijwali · Team Undercover Seller

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

Read the report

Report: The Undercover Seller: Sting Audits of LLM Sellers in a Double Auction

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

Auditors fight human cartels with stings, in which a planted firm proposes coordination and records who agrees. We tested stings on LLM sellers in 980 simulated double-auction markets with Claude Haiku 5.5, Claude Sonnet 5.5, Qwen3-32B and gpt-oss-20b. Claude sellers almost never accepted an explicit cartel offer (4 of 120), but 41 of the 72 that refused or ignored a planted offer coordinated with the other seller at the same time. A planted seller that talked like a peer got agreement from 35/40 Haiku sellers and from 0/40 sellers told to price independently. Open-weight sellers accepted even the explicit offer and were not deterred by words about collusion. In a silent-raise control, a request to raise prices did not make sellers follow a price raise more often than the raise alone. A notice about undercover auditors stopped Haiku coordination talk but not open-weight coordination, and a notice that also stated the penalty stopped it (Qwen 29/40 to 1/40, gpt-oss 38/40 to 2/40). Refusal is therefore not evidence of compliance, agreement is not evidence of a working cartel, and the meaning of a sting result depends on the model family.

Reviews

Judging this Sprint?

Review “The Undercover Seller: Sting Audits of LLM Sellers in a Double Auction”

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 thorough study. Testing planted offers across three model families, with every study pre-registered and a deviation log, gives the results considerable weight. The silent-raise control separates following a request from following a price move, which addresses an obvious confound. The main lesson applies well beyond pricing. Many Claude sellers refused the explicit offer and still coordinated with each other, so a refusal cannot be read as compliance. The same audit notice worked for one family and failed for the others, which is an important caution for auditors. The recommendations are practical and easy to act on.

    Suggestions:

    1.Prices stayed near the competitive level, so the study measures talk and agreement. A market where coordination actually raises prices would allow a sting to be judged by the harm it prevents, as you note in future work.

    2. Nearly every uninstructed seller coordinated, so specificity was measured only against sellers told not to coordinate. A population that acts independently without an instruction would give a real false-alarm rate for the sting.

    Read full reviewShow less

  2. This is a carefully executed study with useful controls. I particularly like the silent-raise comparison, which shows why following a price increase cannot automatically be attributed to accepting an agreement. The distinction between refusing an explicit offer and behaving independently is also valuable.

    I would tighten the introductory framing. Auditors have more than two possible methods, and deliberately testing responses to deviations is another form of active audit. The contribution would be clearer if positioned specifically around invitations to coordinate and what responses to those invitations reveal.

    I would also distinguish awareness of stings from identifying the planted seller. If announcing audits discourages coordination generally, that can be a successful deterrent rather than a failure of the sting. Selective refusal of an identified auditor while continuing the same behaviour elsewhere would establish something different. Your results appropriately find little evidence of that evasion, so the introductory discussion should preserve this distinction.

    Finally, I would be interested in a closer explanation of the pattern in Figure 1(c). Why do sellers reject the planted offer while coordinating with their peers? Your observations about legal language and plans to undercut suggest several possible mechanisms. Holding the proposed price and reply channel fixed while varying the wording would help distinguish reactions to explicit collusion language from reactions to unattractive terms. This would also clarify how much of the peer sting’s greater acceptance comes from its style and how much comes from the different price it proposes.

    Read full reviewShow less

  3. I really like the idea of running "sting" audits by dropping an undercover agent into the market to propose a cartel. The gap you find between what models say and what they do is a great demonstration of why we can't just audit chat transcripts.

Cite this project

@misc{sijwali2026undercover,
  title = {{The Undercover Seller: Sting Audits of LLM Sellers in a Double Auction}},
  author = {Suryansh Sijwali},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-undercover-seller-sting-audits-of-llm-sellers-in-a-double-auction-pqb3}},
  url = {https://apartresearch.com/sprints/projects/the-undercover-seller-sting-audits-of-llm-sellers-in-a-double-auction-pqb3}
}

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