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

Stateful Monitor

Alix Hypher

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

Read the report

Report: Stateful Monitor

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

Current monitors in AI control evaluations are stateless: they re-evaluate each agent action from scratch with no memory of their own prior reasoning. This limits their ability to detect multi-step attacks where malicious behavior is distributed across many individually-innocuous actions. We investigate whether giving the monitor a self-managed scratchpad — a persistent block of notes updated at each step — improves detection in BashArena, a control setting for highly privileged AI agents. We evaluate a scratchpad monitor against a stateless baseline, both using GPT-4.1 mini, on attack and benign trajectories generated by Claude Opus 4.6. We find that neither monitor reliably detects attacks, and the scratchpad monitor performs slightly worse than the stateless baseline. This replicates the negative finding of Kutasov et al. (2025), who observed that persistent notes degraded monitor performance in SHADE-Arena due to anchoring on early benign observations. Our results suggest this limitation generalizes across control settings and is not specific to SHADE-Arena's environment structure. We discuss implications for monitor design and suggest that structured alternatives to free-form note-taking may be needed to make stateful monitoring viable.

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. The results section is unfortunately missing, but the gist is in the Introduction. It's useful to test for generalization of findings like these, makes a lot of sense for a hackathon project. The limitation of not iterating on the scratchpad use strategy is acknowleged, and is the first thing I'd have addressed if I'd had more time. It also would have been nice to see some characterization of the extent to which the provided attacks actually used multi-step strategies where you might expect the scratchpad to be differentially helpful as opposed to mostly single-step attacks.

  2. This is a worthwhile project. However, it's hard to evaluate holistically when the Results section is set to the template instructions. Generally well presented, otherwise, and a relevant replication and extension of the cited paper's results.

Cite this project

@misc{hypher2026stateful,
  title = {{Stateful Monitor}},
  author = {Alix Hypher},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/stateful-monitor-t8dd}},
  url = {https://apartresearch.com/sprints/projects/stateful-monitor-t8dd}
}

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