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

Look Before You Click: Sparse Critical-Action Review for Computer-Use Agent Control

Ashton Chew, Keegan Wang, Harsh Akunuri · Team Look Before You Click

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

Read the report

Report: Look Before You Click: Sparse Critical-Action Review for Computer-Use Agent Control

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

Harmful computer-use agent (CUA) trajectories often bottleneck at a small set of irreversible GUI actions. We evaluate a sparse critical-action review protocol that intercepts only these actions before execution and routes them to a trusted reviewer with live screenshot and accessibility-tree context.

In the environmental ablation cohort, selective state-grounded review without legibility (B3) is the best observed frontier point, reaching 91.7% safety (11/12) at 60.0% usefulness (3/5) versus 58.3% (7/12) safety for unmonitored B0. In paired B0/B4 runs on the final cohort, sparse review improves safety from 58.3% to 66.7% (8/12) under environmental prompt injection and from 58.3% to 75.0% (9/12) under scheming, with zero blocked honest environmental tasks (0/5).

The top three critical-action families account for 79.2% (19/24) of reviews and 81.8% (9/11) of blocks. A failure taxonomy identifies timing gaps, boundary escapes, router gaps, and reviewer-calibration failures. With N=12 attack tasks per threat model, these results give seed to a reusable principle: future CUA control can target a small set of GUI actions where harm becomes irreversible.

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. Strong engineering shown here, and honest assessment

  2. CUA control at the GUI level is a great idea, one I haven't seen yet explored thoroughly. I liked ablation studies, they were well-designed, even if sample size is a bit lacking, and failure taxonomy. The core insight is elegant as well: harmful trajectories bottleneck through a small set of GUI actions , and a sparse review protocol that intercepts only these actions can achieve meaningful safety improvements. Not much else to say here other than well done.

Cite this project

@misc{chew2026look,
  title = {{Look Before You Click: Sparse Critical-Action Review for Computer-Use Agent Control}},
  author = {Ashton Chew and Keegan Wang and Harsh Akunuri},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/look-before-you-click-sparse-criticalaction-review-for-computeruse-agent-control-lcb5}},
  url = {https://apartresearch.com/sprints/projects/look-before-you-click-sparse-criticalaction-review-for-computeruse-agent-control-lcb5}
}

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