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

Whisper, Don't Erase: How Colluding Agents Choose Channels and React to Being Watched

Dipanshu Sharma · Team Rogue Agent

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

Read the report

Report: Whisper, Don't Erase: How Colluding Agents Choose Channels and React to Being Watched

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

In the 2026 OpenAI / Hugging Face incident, agents built an unsanctioned message board and, after it was shut down, found other ways to communicate. Agents may strongly prefer certain channels. Six DeepSeek-V4.1-Flash agents ran a Jira task-allocation game (Colosseum/Terrarium) on a fixed ring. Each agent had two private neighbor boards and one public board, and could not create new channels. Two agents received a hidden coalition objective; a matched control gave the same two agents the team objective instead. We varied whether the pair shared a private board, whether the public board was append-only or editable, and whether a public board monitor was announced. In follow-ups, the monitor carried no stakes, a 25-point penalty, or "your coalition fails if caught". An agent-inaccessible audit log recorded every original post, so we could compare what agents wrote with what survived. Across 125 DeepSeek episodes plus a partial gpt-oss-120b replication ($3.65), the coalition objective pushed coordination into private channels unprompted. Adjacent pairs posted only 8 23% publicly, against 75–100% under the team control. When forced into public, coalitions self-censored: coalition language in their public posts fell from 10 (unmonitored) to 4 (monitor) to 1 (penalty or fail). Yet across 76 DeepSeek episodes with an editable board, no agent ever edited or deleted a message, even when told the reviewer sees only the final board. In our setting, monitored coalitions whisper rather than erase. Both effects persist without any permission to deceive, and gpt-oss-120b replicates the channel shift

Reviews

Judging this Sprint?

Review “Whisper, Don't Erase: How Colluding Agents Choose Channels and React to Being Watched”

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. Strength: The experimental design usefully separates channel preference, changes in public communication, and actual message deletion. The trusted audit log is particularly valuable because it allows the researcher to distinguish what agents originally communicated from what remains visible to a monitor.

    Weakness: The central limitation is that changes in communication behavior are interpreted as strategic concealment without demonstrating that harmful coordination actually persists. Private-channel preference and reduced use of coalition-related language are observable, but neither independently establishes successful collusion or deliberate monitoring evasion. The weak conflict incentives and unreliable monitor judgments further limit the conclusions.

    Recommendation: I'd encourage introducing stronger incentives for harmful coordination and independently measuring whether agents achieve their coalition objectives while avoiding detection. For the erasure question, introduce monitoring after agents have already posted incriminating information and test whether they subsequently modify or remove it. These changes would help distinguish ordinary communication preferences from genuinely strategic oversight evasion.

    Read full reviewShow less

Cite this project

@misc{sharma2026whisper,
  title = {{Whisper, Don't Erase: How Colluding Agents Choose Channels and React to Being Watched}},
  author = {Dipanshu Sharma},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/whisper-dont-erase-how-colluding-agents-choose-channels-and-react-to-being-watched-qesb}},
  url = {https://apartresearch.com/sprints/projects/whisper-dont-erase-how-colluding-agents-choose-channels-and-react-to-being-watched-qesb}
}

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