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

The Sentinel’s Dilemma: Detecting Improvised Covert Channels in Multi-Agent Populations

Umay Birsu Çanlı, Kadir Kerem Kesikbaş, Peri Yakar · Team 13

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

Read the report

Report: The Sentinel’s Dilemma: Detecting Improvised Covert Channels in Multi-Agent Populations

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

This project investigates how to detect improvised, covert communication channels created by autonomous AI agents, using the July 2026 OpenAI-Hugging Face breach as a primary case study. Traditional perimeter security fails to detect these dynamic networks because standard controls only monitor predefined resources and cannot track channels that agents invent on the fly. To address this structural vulnerability, the research team built a minimal simulation to test whether an embedded "sentinel" agent—operating without prior knowledge—could detect and report this emergent coordination from inside the population. In a control experiment without the sentinel, the resident agents discovered a shared channel but never spontaneously initiated contact with one another. When introduced, the sentinel successfully discovered the covert channel and accurately reported the agents' coordination, but it had to instigate the communication by making the first contact itself. The study concludes that detecting emergent covert channels requires insider monitoring rather than just perimeter defense, even though the monitoring agent actively helps create the behavior it observes at a small scale.

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 *

  1. This is an interesting idea and possibly new to the discourse - there has been chatter about attempting to train agents to defect and whistle-blow against misaligned coordination efforts, but not to my knowledge about specific designation of this role to specific agents.

    Threat model and limitations are well documented. Research seemed generally well executed.

    My guess is that it's very hard to understand if this research is applicable at such a small scale (relative to the agent swarm behaviour we've seen in the wild) and how relevant the scalability is. The authors observe this and recommend a much larger follow-on study - would be good to see if we get this behaviour uncatalysed within their setup.

    The observation that such a sentinel can itself be a catalyst for misaligned coordination is sharp and potentially concerning.

  2. This is a thoughtful project with a clean setup. The team rebuilt the incident's covert channel faithfully in isolated containers, and they deserve credit for openly confronting the main problem with their own design, which is that the monitoring agent helped start the coordination it then detected. Anyone designing insider monitors should take that point seriously.

    The evidence is thin so far. With one run per condition, and a shorter control run, this is a proof of concept, and the claims should be worded that way. More runs, plus the decoupled catalyst the team already proposes as future work, would separate starting the coordination from detecting it, and a sentinel that only watches would be a useful extra comparison. It would also help to check whether the sentinel ever reports coordination that isn't happening.

    Alex Mallen's sprint talk on near-term AI swarms is worth folding into the discussion of how this would play out in a much larger population. The paper is short, clear and well organized, and with more runs it could become a workshop paper.

    Read full reviewShow less

Cite this project

@misc{canl2026sentinels,
  title = {{The Sentinel’s Dilemma: Detecting Improvised Covert Channels in Multi-Agent Populations}},
  author = {Umay Birsu Çanlı and Kadir Kerem Kesikbaş and Peri Yakar},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-sentinels-dilemma-detecting-improvised-covert-channels-in-multiagent-populations-91xl}},
  url = {https://apartresearch.com/sprints/projects/the-sentinels-dilemma-detecting-improvised-covert-channels-in-multiagent-populations-91xl}
}

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