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

RecruitmentArena: AI control setting for multi-agent recruitment, manipulation and collusion

Ana Belen Barbero Castejon, Carlos Vecina Tebar · Team RecruitmentArena

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

Read the report

Report: RecruitmentArena: AI control setting for multi-agent recruitment, manipulation and collusion

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

AI control research aims to ensure that misaligned agent behavior is detected or constrained before it causes harm. Existing AI control frameworks focus primarily on single-agent settings and direct misbehavior. But as AI systems progressively operate in multi-agent architectures, a critical gap emerges: harm can arise through inter-agent influence and coordination where no single action reveals the broader intent. We introduce RecruitmentArena, our open-source evaluation framework that treats agent recruitment (one agent inducing others to advance a hidden objective) as a first-class threat in AI control settings.

We contribute (1) a formal definition of recruitment attacks, (2) a six-dimensional taxonomy that characterizes how attacks are initiated, how influence is exerted, how the payload is structured across agents, how recruitment propagates, when it occurs and how it evades detection, and (3) an extensible framework to evaluate recruitment susceptibility, performance-safety trade-off and control protocol effectiveness. Inspired by coordination patterns in online criminal organizations, we instantiate four attack scenarios and evaluate them against two control protocols.

(https://github.com/AnaBelenBarbero/RecruitmentArena)

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. This is an understudied and relevant threat model. It's unclear if your experiments used separate safety and usefulness runs with an attack and honest policy or a single run for each protocol to measure both metrics at once. If you did a single run, I'd be skeptical that your usefulness results were meaningful and of any downstream conclusions.

    In Figure 4, it's unclear what intermediate x-axis values mean. I'd usually interpret this as an average over multiple epochs but this doesn't appear to be the case.

    I don't understand the argument for why measuring number of interventions before detection is a proxy measure of monitoring cost. This does seem like an interesting property of attack strategies, but not very related to monitoring cost.

  2. Agent recruitment is a novel and interesting lens to approach AI Control through. But the recruitment attack taxonomy feels like it's putting the cart before the horse - the dimensions don't seem very well defined and the categories within each seem pretty loose and overlapping, not necessarily cutting reality at the joints. A good qualitative baseline to approach analysis with rather more than a formal taxonomy to run experiments across, and I think this is reflected in that most of the models were unable to complete the task. Overall an excellent foray into analyzing Recruitment as an AI Control surface - the next iteration might benefit from letting observed behavior inform experimental approach more.

Cite this project

@misc{castejon2026recruitmentarena,
  title = {{RecruitmentArena: AI control setting for multi-agent recruitment, manipulation and collusion}},
  author = {Ana Belen Barbero Castejon and Carlos Vecina Tebar},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/recruitmentarena-ai-control-setting-for-multiagent-recruitment-manipulation-and-collusion-5mo2}},
  url = {https://apartresearch.com/sprints/projects/recruitmentarena-ai-control-setting-for-multiagent-recruitment-manipulation-and-collusion-5mo2}
}

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