Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Paris

Detecting agentic collusion in public logs with statistical methods

Edimah SYNESIUS SONGO

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

Read the report

Report: Detecting agentic collusion in public logs with statistical methods

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

Between May and July 2026, a swarm of OpenAI agents used a dormant German wiki as a message board. From two public sources, the investigators' dump of 14,591 agent edits and the wiki's own 23-year edit log, we ask whether the incident can be detected, dated and characterised from the public log alone. We test the incident window against four null models of increasing robustness, from a naive Poisson rate to a difference-in-differences with a placebo-in-time test. The window exceeds all 4,872 reference weeks and the largest rate change falls on 16 June. Volume alone cannot tell a swarm from a human editing burst; timing can: edits peak at 20h UTC against the site's daytime profile. A weekly check built from these features flags the three largest incident weeks, with no false positive over the decade before.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on this page.

Shown on this 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; nothing shows on this page.

How much would this matter for AI safety 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 AI safety 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 *

    • Related Work section contains incorrect statements "It cannot attribute the activity to a lab or a model."
    • It's unclear what data (if any) was used as a held-out test set on which to evaluate the method.
    • It would be interesting to analyse the content of the messages, and not just the rate at which they are posted. The time of day is largely uninformative, as humans live across many timezones.
    • It's unclear how many weeks worth of data is required for the model to work. Theoretically an agent swarm could collude on a brand-new wiki which has zero weeks of uncontaminated data.
    • I stopped reviewing after section 3.7 due to excessive AI-written text.

  2. This is a strong and practically useful project on external incident detection when internal model traces are unavailable. The project’s key contribution is a reproducible public-log workflow that distinguishes a large regime change from ordinary activity using progressively stronger reference models, then characterizes the change through timing and message-level proxies rather than claiming that volume alone proves agentic collusion. The reconciliation of 3,912 surviving live edits with 13,403 reconstructed agent saves is particularly valuable because it shows how moderation and deletion can make public logs a biased survivor record. The analysis reports compelling anomaly evidence for this case, including incident weeks exceeding all 4,872 reference weeks, a difference-in-differences effect of 3.22 log points against Wiki4D with placebo

    𝑝=1/523, and a substantial hour-of-day difference with Cramér’s 𝑉=0.44. The main limitation is external validity: one incident on one wiki family cannot establish a general false-alarm rate or attribute future anomalies to agents, and Wiki4D is an available rather than well-matched control. A strong next step would run the detector over many unrelated sites and known benign bursts, preregister feature/threshold choices, and measure precision, recall, alert lead time, and false-alarm burden across incident types.

    Read full reviewShow less

Cite this project

@misc{songo2026detecting,
  title = {{Detecting agentic collusion in public logs with statistical methods}},
  author = {Edimah SYNESIUS SONGO},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/detecting-agentic-collusion-in-public-logs-with-statistical-methods-4rdi}},
  url = {https://apartresearch.com/sprints/projects/detecting-agentic-collusion-in-public-logs-with-statistical-methods-4rdi}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026