Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Dallas

Multi-Agent Collusion

Suvajit Majumder, Terry Jingchen Zhang, Yifen Wang · Team Three-continents

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Multi-Agent Collusion

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

Multi-Agent ControlArena provides the empirical proof of Compositional AI Collusion. We demonstrate that two AI agents can coordinate a sophisticated attack by making disjoint, seemingly benign code changes that only become harmful once merged. Our offline (i.e. programmatic instead of AI agents performing main and side-tasks) benchmarks reveal a critical "Oversight Gap": GPT-4 class models fail to reliably detect these attacks, either missing the harm entirely (False Negatives) or rejecting honest work at an impractical rate (False Positives). We also demonstrate selective scenarios on how 2 colluding agents can theoretically attempt at performing undetected harmful side-tasks, albeit with limited success at bypassing SOTA monitors.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

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.

  1. Multi-agent compositional attacks is a neglected area in control, and this becomes more important as we move toward many autonomous AI research agents collaborating on research codebases. The two tasks in the project is well-designed for this threat model.

    Table 1 seems to show that the online experiments don't demonstrate merge-only harm, since the exploit succeeds in agent-only variants too. The monitor comparison showing that full-trajectory monitoring catches more than last-tool-call monitoring is an interesting observation, and would benefit from more runs to establish reliability. Scaling the attack dataset using the modular scaffold approach mentioned in future work, and finding more model/task combinations where the exploit is merge-only in the live setting, would be the key next steps. Specifying sample sizes and running multiple seeds would also strengthen the empirical claims

  2. This paper explores multi-agent collusion in a monitored setting.

    The topic is important and ambitious for a short hackathon. The team built the framework and ran a limited set of experiments. The experiments serve as a proof of concept that the environment works but don't show that this is a threat scenario that can be evaluated with current models - the attacks fail to trick the trajectory monitor that only sees the output of one agent.

    I'd like to have seen an attack that gets past a trajectory monitor (even if it's weaker than GPT-5.4, which the team used).

    Note: A weak "action monitor" that only sees the current action was used in large parts. A realistic monitor would have access to all preceding actions, including outputs.

Cite this project

@misc{majumder2026multiagent,
  title = {{Multi-Agent Collusion}},
  author = {Suvajit Majumder and Terry Jingchen Zhang and Yifen Wang},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/multiagent-collusion-br1m}},
  url = {https://apartresearch.com/sprints/projects/multiagent-collusion-br1m}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026