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

BCP-1: Proof-Carrying Containment for Guardrails-Off AI Evaluations

Khadija Edarzi · Team ZKL

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

Read the report

Report: BCP-1: Proof-Carrying Containment for Guardrails-Off AI Evaluations

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

A sandbox can keep an allowed network path, raise detections, and still fail to contain. In the July 2026 OpenAI–Hugging Face incident the model used reachable package infrastructure; detections did not force isolation.

BCP-1 is an eight-control containment profile for dangerous-capability evaluations with production safeguards off. This project specifies those controls and implements the authorization and evidence slice in Python: signed manifest, one-use kill-epoch capabilities, credential broker, bounded evidence, streaming verifier, and supervisor receipt. The local isolate is signed SIMULATED; it is not a production microVM.

On that reference, five selected control omissions each produced a non-zero unsafe count; the full contract scored zero. Access logs verified 0/5 predefined claims, signed-line logs 1/5, and the BCP-1 bundle 5/5 on the mediated path (4 events vs 50,000 records). Four scenarios and 32 tests passed.

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 core architecture presented here is interesting but repeats in a several submissions I saw for this sprint. I particularly liked your approach to permit lifetime management and auditability.

    A lot of the ideas you present here where remind me of existing frameworks (certificate transparency logs cosigning comes to mind.

    The writing itself was clear, while a bit dense. I think the "proof-carrying" framing is a bit misleading here, what you get is more of an attestation to the claim. Some of your citations carry the wrong date or misquote the title of the cited work.

  2. Binding actions to signed permissions is a strong way to stop a model causing real-world effects it was never authorised for. However the manifest has to name allowed destinations up front, and evals are exactly where you can't predict what the model will reach for, so unanticipated but harmless actions become denials needing a human. The purpose of evals is also to see where the model deviates, and not all deviations are harmful.

Cite this project

@misc{edarzi2026bcp1,
  title = {{BCP-1: Proof-Carrying Containment for Guardrails-Off AI Evaluations}},
  author = {Khadija Edarzi},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/bcp1-proofcarrying-containment-for-guardrailsoff-ai-evaluations-opff}},
  url = {https://apartresearch.com/sprints/projects/bcp1-proofcarrying-containment-for-guardrailsoff-ai-evaluations-opff}
}

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