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

Trigger Happy: Trusted Monitors Mistake Noisy Failures for Adversarial Intent

Prithviraj Singh Shahani, Ayush Jain · Team MoxFood

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

Read the report

Report: Trigger Happy: Trusted Monitors Mistake Noisy Failures for Adversarial Intent

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

We set out to test whether LLM-based monitors can detect adversarial intent when the harmful actions themselves fail, and whether that detection holds up when the evidence of failure is removed. We built two LinuxArena environments where OS permissions make the adversarial side task impossible, ran attack-policy agents against them, and pruned failure steps from the resulting trajectories. We found that monitors over-rely on the noise of failed actions: removing them drops safety curve performance and shifts monitor narratives from confident accusations to hedged suspicion.

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. Excellent framing and addressing of a problem which is not yet well known, but nonetheless has the potential to significantly impact real-world validity and deployment. While the novelty of this seems quite substantial, there are a number of areas where improved methodological steps would meaningfully increase the overall value:

    (1) Experiments were conducted on a very small sample size. With a change in scores as small as it is, and the main difference being qualitative assessment, the work simply needs more runs to have legs.

    (2) A simple but deterministic rule for which steps to remove would have been an improvement over "asking an LLM to do it," especially when you didn't have sample size on your side.

  2. I like this! The overall premise is subtle and significant, in that detection is confounded by models overindexing on failure vs success. this is quite plausible in an RL setting. I think the setup makes sense and the premise is decently explained. There aren't many tests, but given the short timespan I think it's fine to prioritise quality over quantity.

Cite this project

@misc{shahani2026trigger,
  title = {{Trigger Happy: Trusted Monitors Mistake Noisy Failures for Adversarial Intent}},
  author = {Prithviraj Singh Shahani and Ayush Jain},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/trigger-happy-trusted-monitors-mistake-noisy-failures-for-adversarial-intent-xeoz}},
  url = {https://apartresearch.com/sprints/projects/trigger-happy-trusted-monitors-mistake-noisy-failures-for-adversarial-intent-xeoz}
}

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