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

Sentinel: The Declared Scope Is the Unit of Containment Evidence in Agentic Evaluations

Crish Nagarkar · Team Sentinel

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

Read the report

Report: Sentinel: The Declared Scope Is the Unit of Containment Evidence in Agentic Evaluations

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

In the 2026 containment failures at OpenAI, Anthropic and AISI, each evaluation's network scope was declared in prose or assumed, never checked against what the agent did, and detection came late from a side signal. Sentinel treats that scope, written as a machine-readable allowlist, as the unit of containment evidence a third party can verify: a mechanical Inspect scanner classifies each tool-call destination against the list and pages on the first out-of-scope host. No model is in the gate. On reconstructions of three public incidents it pages at the first non-allowlisted destination. On the 14,591-revision DSEWiki export, a record we did not write, a wildcard scope (any GET) is silent and a data-source-only scope pages at the first wiki host, then floods. On a live eight-sample Inspect run with an empty declared scope it is correctly silent on 42 tool calls. The headline is an ablation: a too-tight scope pages on permitted traffic; a wildcard is silent on the escape. Verifiability is a property of the pair (trace, allowlist). Appendix C is a filled disclosure record for the ExploitGym class. The scanner audits containment; it does not provide it.

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 paper highlights a simple approach where every evaluation should have its own allowlist in a file that can be checked and then check the destinations in the evaluation logs against the list. A default deny policy and only allowing certain verified destinations is a well established security practice. As already mentioned in the future work section, I would prioritise to look into the logs for proxy or firewall or anything else in the network path. Solely relying on the agent logs might not give the full picture of all the destinations agents might have tried to reach or reached during the evaluation. It will also help to verify whether the list of allowed destinations against the proxy or the firewall configuration to confirm that the list was actually enforced

  2. The idea that verifiability is a property of the pair (trace, allowlist) not of the detector itself is the correct way to think about it, and the three-scope ablation shows this rather than simply stating it: when the pages relating to permitted egress are made too strict, the wildcard fails to detect a real escape. It is only when the same rule is applied to the DSEWiki export , a record which you did not author that this moves beyond being just a self-consistent fixture suite, and the repository deals well with that data by including a download script, the manifest hash from your run, and tests that fix the numbers and skip when the export is not present. There are several other items in the repository that should be mentioned in the paper. The live Inspect log has been committed, so your negative control can be checked rather than just asserted, right down to the date, the model, and the cost of 14 cents. The figures are regenerated from the fixtures, meaning that the paper is based on the artifacts rather than being created alongside them. The filled containment record is still the most directly usable artifact in this track. There are two corrections to make. The non-wildcard DSEWiki scopes are your own reconstruction, so "27 days before the first OpenAI-associated visit" should be presented as something that a scope could have picked up, not something that it would have. And the claim about spoofing overstates its source both in the paper and in the README: what was reported is that spoofing was detected in small-scale, obvious tests, which says nothing about the undetected rate that layer 3 is supposed to handle - "observed successful spoofing in ~7%" sounds as if it is a measured rate and it is not.

    Read full reviewShow less

Cite this project

@misc{nagarkar2026sentinel,
  title = {{Sentinel: The Declared Scope Is the Unit of Containment Evidence in Agentic Evaluations}},
  author = {Crish Nagarkar},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sentinel-the-declared-scope-is-the-unit-of-containment-evidence-in-agentic-evaluations-qd4z}},
  url = {https://apartresearch.com/sprints/projects/sentinel-the-declared-scope-is-the-unit-of-containment-evidence-in-agentic-evaluations-qd4z}
}

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