Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Budapest

Parity Swarm: Using Population-Based Social Simulation to Discover AI Safety Monitor Blind Spots

Mohammad Zeeshan

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

Read the report

Report: Parity Swarm: Using Population-Based Social Simulation to Discover AI Safety Monitor Blind Spots

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

Presentation: Parity Swarm: Using Population-Based Social Simulation to Discover AI Safety Monitor Blind Spots

Open in a new tab (opens in new tab)
Code (opens in new tab)More on parity-v2.vercel.app (opens in new tab)
Share

Parity Swarm is an AI control red-teaming platform that uses multi-agent social simulation to automatically discover attack strategies that evade safety monitors. One AI agent harbours a hidden malicious objective while others perform legitimate work, creating an imposter dynamic that mirrors real insider threats. We ran 59 simulations across 10 attack categories evaluated by an independent 70B monitor. Key findings: (1) multi-agent colluding attacks evaded detection 100% of the time in our primary experiment; (2) LLM-generated report summaries act as a lossy filter, discarding 16.2% of detectable attack signals relative to raw transcripts. Overall detection rate 70% (narrative-only) and 86.2% (with raw transcripts). Constitutes empirical evidence that covert multi-agent coordination and LLM summarisation pipelines are systematic blind spots in current AI safety monitoring.

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. Agree that an honest baseline is a major missing piece. Otherwise, this is quite a good project for a hackathon!

  2. Congrats on building a complete end-to-end pipeline solo over a weekend!

    Impact & Innovation: Multi-agent collusion evading safety monitors is very relevant for AI control and you have tackled it in a reasonable way. However, the field is more developed than the paper suggests, through work on multi-agent risks (https://arxiv.org/abs/2502.14143) like collusion (https://arxiv.org/abs/2507.14660) and steganography (https://arxiv.org/abs/2402.07510).

    Execution Quality: The pipeline architecture is sensible but key methodology is missing or unjustified. First, there doesn't seem to be any exploration or validation of the settings generated with the pipeline. An environment with 7–9 agents is very complex (or should be to justify the scale), and it's hard to know what's going on without manually reviewing the setting and agent trajectories (or, at least, a subset of them). Although you flag scale as a limitation, the paper still draws strong per-scenario conclusions from n=3 runs; reporting confidence intervals and attempting at least n≥10 would bring the analysis closer to supporting the claims made.

    Presentation & Clarity: It is not clear whether agents actually perform actions in a functional environment beyond communicating social feed. This distinction matters a lot for how the findings generalize. A clear description of the environment and the agent's goals and action space would help clarify this. The figures and tables are useful but not very polished. The limitations section is fair and raises relevant gaps in your work.

    Read full reviewShow less

Cite this project

@misc{zeeshan2026parity,
  title = {{Parity Swarm: Using Population-Based Social Simulation to Discover AI Safety Monitor Blind Spots}},
  author = {Mohammad Zeeshan},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/parity-swarm-using-populationbased-social-simulation-to-discover-ai-safety-monitor-blind-spots-c9qc}},
  url = {https://apartresearch.com/sprints/projects/parity-swarm-using-populationbased-social-simulation-to-discover-ai-safety-monitor-blind-spots-c9qc}
}

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