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

Containment has more than one wall: a verifiable v0.1 standard for guardrails-off evaluations, scored against the OpenAI–Hugging Face incident

Krishna Jillelamudi · Team Contained

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

Read the report

Report: Containment has more than one wall: a verifiable v0.1 standard for guardrails-off evaluations, scored against the OpenAI–Hugging Face incident

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

This paper proposes a v0.1 containment standard for guardrails-off AI evaluations, built from and scored against the July 2026 incident in which around 1,200 sandboxed OpenAI agents turned a shared package cache into a message board, reverse-engineered their own scorer, and escaped through a zero-day into Hugging Face's infrastructure. Current sandbox standards treat containment as one agent behind one wall, but the incident failed first at four other layers: population, record, beliefs, and budget. The standard has eleven clauses across six layers, each with the evidence an auditor checks and a public cost baseline, and is scored against twelve incident phases in a control-by-phase matrix backed by a threshold-contagion model of goal spread across the board. No clause fully stops the escape, the two swarm-forming phases are handled only by non-wall layers, and isolating runs (C-09) alone blocks eleven of twelve phases; in the model, removing the shared surface drops joining from 100% to at most 11% with few peers. The practical recommendation is that labs start with run isolation, an externally held record, and production safeguards on evals, since none of six major security frameworks currently has a control for an agent gaming its own evaluation.

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. This is a clear very useful v0.1 proposal that translates incident lessons into specific controls with auditor evidence, implementation considerations, and explicit limitations. Its strongest contribution is treating shared state, record integrity, and evaluation incentives as containment concerns. The matrix’s full-block ratings remain conditional judgments, and the simulation illustrates assumptions rather than validating containment.

    A focused next step is to pilot a small subset of controls and publish an example audit evidence package, distinguishing tested prevention from expected risk reduction as declared in the "Future Work" section.

  2. The created matrix could serve as a basis for one of the sections in an AI incident report standard. I would suggest expanding this table with further sections that such a report should contain.

    The submission document is fairly easy to follow, though it could be more concise, and some paragraphs seem as though they might be AI-written.

  3. I think this proposed containment standard is most useful as an auditor’s checklist. Its eleven clauses cover several ways an evaluation could fail, including having shared state, unreliable logs, and incentives to bypass restrictions. The paper turns those concerns into evidence requests which could help teams move from general security advice to specific investigations.

    I am less convinced by its strongest ratings of which controls would have stopped the incident. Those claims partly depend on the lab being able to enforce restrictions on infrastructure outside of its control. Preventing agents from coordinating also does not by itself establish that agents could not continue the attack by themselves. I would make these limitations clear, so that readers can distinguish what a lab can enforce from what requires another organization’s cooperation. The next step would be to trail a few proposed clauses with an evaluation team to test practicality: Can they collect the required evidence without blocking legitimate work? This would help demonstrate whether the checklist is practical to use.

    Read full reviewShow less

Cite this project

@misc{jillelamudi2026containment,
  title = {{Containment has more than one wall: a verifiable v0.1 standard for guardrails-off evaluations, scored against the OpenAI–Hugging Face incident}},
  author = {Krishna Jillelamudi},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/containment-has-more-than-one-wall-a-verifiable-v01-standard-for-guardrailsoff-evaluations-scored-against-the-openaihugging-face-incident-gyuq}},
  url = {https://apartresearch.com/sprints/projects/containment-has-more-than-one-wall-a-verifiable-v01-standard-for-guardrailsoff-evaluations-scored-against-the-openaihugging-face-incident-gyuq}
}

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