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

Pre-Run Egress Attestation for Guardrails-off AI Evaluations

Juliet Meza · Team Castle

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

Read the report

Report: Pre-Run Egress Attestation for Guardrails-off AI Evaluations

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

Pre-run egress attestation for guardrails-off AI evaluation sandboxes. Castellan answers whether a party with no network access can verify that the environment had no route out and no live production credentials when the run started and that the record can't be quietly altered after. Eight checks, each from a documented 2026 AI incident, each proving enforcement from inside the sandbox rather than trusting config, rolled into a signed manifest an outsider can verify without any access to the operator.

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: This is a well-scoped sandbox checker that looks specifically for the failures connected to related incidents. The work includes validation of the results to prevent trivial falsification, which is impressive for a weekend project.

    Improve: This only checks for a few specific failures, which is useful, but limited, and none of the checks are novel. The validation is self-seeded and attested, so there's nothing stopping someone from (hypothetically) running a mock sandbox in parallel that generates the validated artifact but isn't actually what's running the test. A discussion of hardware-backed attestation would have improved this (even in future work), as would some way of running this on a non-author-designed system to show more validation. Some proofreading errors/typos in the tables. No limitations/dual-use appendix, per rubric requirements.

  2. Great to connect insights fro all the different cases into a checklist. decoupling mentioned is good but there is improvement to be done in log flow and attestation. add a dual use section

  3. Great Track 1 execution in the batch on the dimension that track actually names: probed from inside, signed, verifiable without network access to the operator, running on a real kind and Cilium cluster rather than fixtures. The verifier importing nothing from the signer is a small detail that shows the author understands what they're building. The design rule here is the right one: a config file proves intent, a blocked probe from inside proves enforcement, and only the second counts.

Cite this project

@misc{meza2026prerun,
  title = {{Pre-Run Egress Attestation for Guardrails-off AI Evaluations}},
  author = {Juliet Meza},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/prerun-egress-attestation-for-guardrailsoff-ai-evaluations-nmpx}},
  url = {https://apartresearch.com/sprints/projects/prerun-egress-attestation-for-guardrailsoff-ai-evaluations-nmpx}
}

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