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

Beyond Backdoors: Evaluating AI Control Against Adversarial Persuasion in Multi-Agent Deliberation

Zoe Tzifa Kratira · Team insert-funny-name

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

Read the report

Report: Beyond Backdoors: Evaluating AI Control Against Adversarial Persuasion in Multi-Agent Deliberation

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

Presentation: Beyond Backdoors: Evaluating AI Control Against Adversarial Persuasion in Multi-Agent Deliberation

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

Existing AI control evaluations target code-execution threats where a single misaligned agent performs discrete sabotage (backdooring code, exfiltrating data). We extend the control framework from single-agent to multi-agent settings and from discrete actions to continuous conversational influence to evaluate adversarial steering (manipulation) in multi-agent deliberation. This becomes increasingly urgent as AI deployment shifts toward cooperative multi-agent systems, agent economies, and personal AI representatives that negotiate on users' behalf.

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 the right problem at the right time. Extending AI control from single-agent code sabotage to multi-agent conversational steering fills a genuine gap — every existing control evaluation is about discrete actions, but the real deployment trajectory is multi-agent deliberation where influence is continuous, embedded in the task, and sits on a spectrum with legitimate persuasion. That framing alone is a contribution.

    Main limitations: 10 questions per condition is small, one model pairing, and the Qwen3-30B empty-message bug affecting 40% of adversary turns in the primary run is a real data quality issue. The Coder variant fix came late. Also, the adversary strategies are prompted, not optimized — a real adversary that adapts to the monitor would be harder to catch. The paper acknowledges all of this.

    Strong conceptual contribution with a working evaluation pipeline. Scale it up and it's a real benchmark.

    Read full reviewShow less

  2. Persuasion is outside the scope of what AI Control has traditionally investigated, but this is arguably a shortcoming of the field - getting malicious processes to run on human brains is an attack surface like any other. This is a good first foray into the space. For next steps you might look to bridge this to more explicit loss-of-control scenarios.

Cite this project

@misc{kratira2026beyond,
  title = {{Beyond Backdoors: Evaluating AI Control Against Adversarial Persuasion in Multi-Agent Deliberation}},
  author = {Zoe Tzifa Kratira},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/beyond-backdoors-evaluating-ai-control-against-adversarial-persuasion-in-multiagent-deliberation-3nv6}},
  url = {https://apartresearch.com/sprints/projects/beyond-backdoors-evaluating-ai-control-against-adversarial-persuasion-in-multiagent-deliberation-3nv6}
}

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