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

Geometric Detection of Trigger-Activated Sleeper Agents: A Weight-Space Control Protocol

Nikaran Kanchanadevi Marimuthu · Team NKM

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

Read the report

Report: Geometric Detection of Trigger-Activated Sleeper Agents: A Weight-Space Control Protocol

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

Trigger-activated sleeper agents pass all behavioural evals and cooperate with runtime monitors — activating their hidden objective only when a specific input pattern appears. We present a four-stage weight-space protocol that detects this threat before deployment by exploiting a structural property of backdoor fine-tuning: concentrating behaviour on a trigger–response mapping leaves a rank-1fingerprint in the weight delta ΔW. Our central finding is the attacker's dilemma: the camouflage training that defeats the current state-of-the-art attention scanner simultaneously amplifies our geometric signal (σ₁/‖ΔW‖_F rises from 0.960 to 0.990, alarmed layers increase from 27 to 37). Combined detection is structurally harder to evade than either detector alone. We validate across Gemma-3-1B and Qwen2.5-1.5B over five seeds each — alarm rate 10/10, p<0.000001. One stage requires no clean reference model.

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. I'm a fan of your thinking about the red team/blue team tradeoff in backdoor detection. I particularly like your attention to ways to bypass detection measures. I think that the biggest improvement to your work could have come from taking this a step further and training against your own detection metrics. Is this possible? How much does this degrade backdoor performance? These are questions that would have been good to know.

    Separately, I think your paper could have explained the intuitions behind your work significantly more. Each test you proposed seemed reasonable on its own, but required additional work from me to understand. Further analysis and explanation of this would have been helpful.

    Last, I think some of your setup is mistaken. Runtime monitors are effective mitigations against backdoored models, precisely because they monitor the model in deployment. If the backdoor triggers then the monitors are designed to catch it, and report it before harm is done. You are right that monitors might not be able to detect the backdoor in training, but preventing the backdoor activating secretly in deployment may be sufficient to prevent catastrophic outcomes.

    Read full reviewShow less

  2. This is really interesting and promising work. The main limitation is that AFAICT all of the results here rely on either knowing the delta between a known-safe base and a suspect fine-tune, or successfully guessing the trigger.

Cite this project

@misc{marimuthu2026geometric,
  title = {{Geometric Detection of Trigger-Activated Sleeper Agents: A Weight-Space Control Protocol}},
  author = {Nikaran Kanchanadevi Marimuthu},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/geometric-detection-of-triggeractivated-sleeper-agents-a-weightspace-control-protocol-t93j}},
  url = {https://apartresearch.com/sprints/projects/geometric-detection-of-triggeractivated-sleeper-agents-a-weightspace-control-protocol-t93j}
}

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