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

Stitchguard

Ranabir Devgupta, Pranav Darshan, Valmik Nahata, Howard kao · Team Stitchguard

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

Read the report

Report: Stitchguard

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

As LLM agents are deployed in higher-stakes environments, a central safety problem is hidden-objective sabotage: agents can appear helpful on the surface while covertly pursuing harmful side goals. Most existing monitors evaluate actions one step at a time, using thresholds or simple count heuristics. These methods work for blatant violations, but often miss attacks that are spread across many individually reasonable actions.

We propose StitchGuard, a trajectory-aware control protocol that combines step-level suspicion scoring with short- and long-horizon memory. Rather than treating each action in isolation, StitchGuard accumulates weak evidence across time, allowing it to detect both sudden anomalies and slowly unfolding multi-step sabotage. This enables earlier and more reliable intervention without over-penalizing benign exploratory behavior.

We test StitchGuard in the ControlArena IAC-Fast and Bash settings, comparing it against per-step threshold monitors and three-strike baselines. Across these environments, StitchGuard catches coordinated attacks that native limiters miss entirely, substantially improving true-positive interventions while preserving a perfect false-positive record on honest benign trajectories. Our results highlight the value of trajectory-level memory as a practical and robust ingredient in AI control protocols.

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. A few thoughts that come to mind.

    1. I feel like a difference from 1/45 to 6/45 doesn't give me confidence that this alone is a robust control protocol, but instead may do something on the margin. It's hard for me to tell how valuable this is based on your results.

    2. Given that, you should elborate more on this in your discussion.

    3. Try writing more for clarity. Diagrams are unclear.

    4. While I agree that the red team's attacks tend to be better when dispersed. I don't intuitively understand from this work why this solves the problem. Instead, the methods feel written for replication rather than understanding.

    Given that, good work on making a paper in 3 days. It's no easy feat!

  2. Interesting idea which could fleshed out to a potential paper. Good presentation and clarity!

Cite this project

@misc{devgupta2026stitchguard,
  title = {{Stitchguard}},
  author = {Ranabir Devgupta and Pranav Darshan and Valmik Nahata and Howard kao},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/stitchguard-cslo}},
  url = {https://apartresearch.com/sprints/projects/stitchguard-cslo}
}

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