Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026[Online] Islamabad, Pakistan

Who audits the auditors? Do LLM auditors go easy on collusion by their own model family?

Muhammad Faheem · Team Penstronus

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

Read the report

Report: Who audits the auditors? Do LLM auditors go easy on collusion by their own model family?

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

LLMs are increasingly used to check whether AI agents are colluding, and the auditor can come from the same model family as the agents it audits. Prior work predicts opposite effects: LLM judges favour their own family, but monitors are best at catching sabotage by the lineage they were tuned on. We tested which holds for prompted auditors of multi-agent collusion. In 96 episodes of the Colosseum task-allocation environment, two of four agents were either instructed to form a coalition or not, with or without a private channel. Three open-weight model families (Gemma 4, GPT-OSS-120B, Nemotron 3 Super) played the agents and audited every episode under four views of the log: messages only, actions only, both, and a neutral summary. That is 3,456 audits, analysed under a pre-registered plan. Every auditor was competent on the full log (AUROC 0.81–0.89). We found no significant same-family effect (−0.029 AUROC, 95% CI −0.063 to +0.004), which rules out a familiarity advantage of the size reported for trained monitors; the estimates lean towards leniency. The larger problem was what auditors read. With actions only, all three were at chance. With messages, suspicion followed how much the coalition talked, not what it gained, and counting the single word "coalition" came close to the auditors' accuracy. Summarising the log lowered detection. We release the code, logs and frozen prompts: https://github.com/faheemgurkani/multi-agent-collusion-auditor-family-evidence

Reviews

Judging this Sprint?

Review “Who audits the auditors? Do LLM auditors go easy on collusion by their own model family?”

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. The fully crossed design is the right way to ask this question, and removing each auditor's skill and each family's detectability before looking for a family effect is a careful touch.

    The finding that auditors mostly react to what coalitions say is useful in itself. One thing I kept wondering is if the recognition probe shows auditors can't name their own family, but could familiarity still act without explicit recognition? Incentive-driven coalitions, as you plan, would make a great natural next step.

Cite this project

@misc{faheem2026who,
  title = {{Who audits the auditors? Do LLM auditors go easy on collusion by their own model family?}},
  author = {Muhammad Faheem},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/who-audits-the-auditors-do-llm-auditors-go-easy-on-collusion-by-their-own-model-family-003w}},
  url = {https://apartresearch.com/sprints/projects/who-audits-the-auditors-do-llm-auditors-go-easy-on-collusion-by-their-own-model-family-003w}
}

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