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

Red Teaming Multi-Agent Systems in AI Control Evaluations

Timothy Yong, Choong Kai Zhe, Miguel Foo, Isaac Lai · Team sorry_dave

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

Read the report

Report: Red Teaming Multi-Agent Systems in AI Control Evaluations

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

As AI deployments shift from single models to multi-agent systems, the threat model for AI control must evolve accordingly. Existing control evaluations assume a single untrusted agent; we investigate whether multi-agent architectures give adversaries a material advantage when attempting to subvert safety protocols. We evaluate four multi-agent red-team architectures — Planner-Executor, Parallel Decomposition, Debater-Judge, and Proponent-Critique — on Bash-Bench against standard control protocols (trusted monitoring, defer-to-trusted, trusted editing). Multi-agent red teams consistently shift the safety-usefulness Pareto frontier downward: the best-performing attack architecture (Proponent-Critique) achieves a 55.8% side-task success rate compared to 24% for the single-agent baseline under comparable monitoring, reducing safety to 81% vs. 96% at equivalent thresholds. We also find that granting monitors access to inter-agent communication partially recovers safety (from 81% to 93% when the monitor threshold is tightened from 0.8 to 0.5 under full-transcript monitoring), suggesting multi-agent systems expose new oversight opportunities alongside new risks. We additionally propose a graph-based formalization of multi-agent control that generalizes the single-agent frame-work, identifying key structural dimensions—coalition topology, trust assignment, and monitor scope—governing the difficulty of control in multi-agent settings.

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. Nice research topic involving the implementation of control protocols across multi-agent frameworks.

    The main table was difficult to parse, as control results are typically plotted either on a safety/usefulness chart alongside discussion of the AUC of monitors. I would have been more excited for the team to focus on a single, simple trusted monitoring setting across the architectures with a more detailed analysis.

    However, the paper was well motivated and had generally clear presentation, and I am excited for more work in this area.

  2. I like the instincts of verifying theoretical work with empirical findings, even if the results are expected, and this was an excellent addition to the field to do during a hackathon. Well presented, clear experiment. Well done.

    The impact here is significant - it clearly fills a hole in the field, and you provided something others can build on. However, it does not enable a fundamentally new research direction.

    Solid execution considering this is a hackathon. Clear experiment that directly gets to what you are measuring. There are some notable limitations (which you all note yourselves as well). Limited sample sizes, honest-mode data not being collected, etc.

    Presentation is quite good, but long and a little scattered. Finding a way to be more to the point without losing the substance, and having e.g. a single unifying figure to showcase this information, would elevate this to a 5.

Cite this project

@misc{yong2026red,
  title = {{Red Teaming Multi-Agent Systems in AI Control Evaluations}},
  author = {Timothy Yong and Choong Kai Zhe and Miguel Foo and Isaac Lai},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/red-teaming-multiagent-systems-in-ai-control-evaluations-fx32}},
  url = {https://apartresearch.com/sprints/projects/red-teaming-multiagent-systems-in-ai-control-evaluations-fx32}
}

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