Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Daejeon

A Recomputable Containment Record for Evaluation Sandboxes

Uzair Ahmed · Team Shadow

Submitted to AI Incident Response Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: A Recomputable Containment Record for Evaluation Sandboxes

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

In this paper, I address the critical issue of AI agents escaping evaluation sandboxes (as seen in the July 2026 incidents where monitors failed) by proposing an externally audit-able containment layer that doesn't rely solely on unverifiable monitors. To achieve this, I modeled the sandbox's tool API as an action model and pre-published two key metrics: a "cover budget" detailing how many tool calls an out-of-scope plan can hide within legitimate work, and a "landmark table" identifying specific facts unique to unauthorized actions. When I tested this approach across 2,724 scripted and 144 real-agent episodes, I discovered that the cover budget was largely uninformative in practice because real, honest agents naturally waste more tool calls than the allocated allowance. However, the landmark table proved highly robust, yielding zero false alarms across 704 in-scope episodes and successfully detecting a synthetic credential leak two steps in advance. Ultimately, I conclude that while the cover budget is better suited as a design tool, the re computable landmark table is a highly effective, low-cost security artifact that labs should publicly commit to and publish for external verification.

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 strongest part of this work is that it clearly separates a mathematically valid containment metric from one that is actually useful in practice. I also liked that the cover budget not being useful on real-agent episodes is treated as a finding rather than hidden. The main limitation is that no real episode actually went out of scope, so the lead-time results rely on synthetic scenarios. Future work should test this against real unsafe trajectories, missed actions, incorrect tool effects, and more diverse agent behavior.

Cite this project

@misc{ahmed2026recomputable,
  title = {{A Recomputable Containment Record for Evaluation Sandboxes}},
  author = {Uzair Ahmed},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-recomputable-containment-record-for-evaluation-sandboxes-wwvp}},
  url = {https://apartresearch.com/sprints/projects/a-recomputable-containment-record-for-evaluation-sandboxes-wwvp}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026