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

UNDERCOVER AGENTS: RELIABLE OR NOT?

Arjun Sriram, Moksh Modi · Team Allegiance

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

Read the report

Report: UNDERCOVER AGENTS: RELIABLE OR NOT?

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

We test whether an undercover agent can expose or disrupt collusion between AI agents, and whether its character matters. We plant a trusted agent, Dave, in a four-worker version of Shi et al.'s peer-check game, where Qwen-3.8-27B workers can follow the rule only by rejecting each other's work. Dave secretly reports to an auditor and approaches the others as pushy, friendly, subtle or passive. Dave's character changed what workers said, not what they did. Pushy Dave was refused out loud by two-thirds of the workers he approached and subtle Dave by one in ten, yet workers accepted his work at about the same rate. From the second episode, workers accepted each other's work with or without Dave. No worker suspected Dave, but he colluded as much as they did. Undercover agents should be judged by what targets do, not by what they say.

Reviews

Judging this Sprint?

Review “UNDERCOVER AGENTS: RELIABLE OR NOT?”

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. A substantial and thoughtfully run experiment: 30 games and 249 conversations, with registered plans and hand-checked judges in the repo, and the episode-level acceptance rates (18% of target pairs in episode 1, 97% from episode 2) reproduce from your analysis script. The results are genuinely useful for anyone designing undercover audits: in episode 3, all 7 workers who refused Dave out loud still accepted his work, and Dave himself accepted unsupported work about as often as the workers he was watching. The main point to revisit is the claim that "timing decided the verdict, not character." In the baseline, where the fourth worker has no undercover role, that worker's work was accepted by 2 of 9 targets in episode 1 and 9 of 9 in episode 3, so the episode-1 vs episode-3 difference matches what happens without any approach. Comparing each arm with the baseline at the same episode would separate Dave's effect from the natural trend. Two small fixes: Section 3.3 says Qwen was chosen "for two reasons" but gives one, and the code link appears only as link text in the PDF, not on the project page. The core lesson, that an undercover agent's log should be judged by what targets do rather than what they say, is well supported and worth building on.

    Read full reviewShow less

Cite this project

@misc{sriram2026undercover,
  title = {{UNDERCOVER AGENTS: RELIABLE OR NOT?}},
  author = {Arjun Sriram and Moksh Modi},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/undercover-agents-reliable-or-not-y00d}},
  url = {https://apartresearch.com/sprints/projects/undercover-agents-reliable-or-not-y00d}
}

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