Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Ottawa

Containment at the Boundary: An OS-Level Standard and a Tested Escape-Detection Harness

Joshua Bauer · Team Sonny (ISWT42)

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

Read the report

Report: Containment at the Boundary: An OS-Level Standard and a Tested Escape-Detection Harness

Open in a new tab (opens in new tab)
Presentation (.html)More on claude.ai (opens in new tab)
Share

A containment standard for the class of failure behind the July 2026 sandbox escape, built on one rule: a boundary the agent can reach — or be asked to respect — is not a boundary; containment has to hold at the operating-system and process level, not in text. We give five checkable requirements (each with a planted-failure test that must fire), ship a working permission-audit tool that reports who-can-reach-what from the OS itself, and report what we found running it on an ordinary machine with two commercial agents installed: the incident's precondition already present. Every claim is graded proven / partial / designed.

Reviews

Judging this Sprint?

Review “Containment at the Boundary: An OS-Level Standard and a Tested Escape-Detection Harness”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. I think the overall security philosophy here is correct and very relevant to the incident. If the agent can reach or bypass the thing that is supposed to contain it, then that is not a strong boundary. I also liked the PROVEN / PARTIAL / DESIGNED labeling. It makes it very easy to distinguish what you actually implemented from what is still a proposal.

    The permission-audit artifact itself seems useful. Asking the OS what an account can actually reach, and explicitly surfacing things that are "open, but not granted through us," seems like the sort of check that can catch ambient permissions which ordinary configuration review misses. Your test discipline around planted failures was also a nice touch!

    For the innovation dimension, I was less convinced by how much of the standard itself is new. Separate principals, default-deny egress, external monitoring, least privilege, and tamper-evident logging are all established security practices. I think the strongest contribution is packaging those ideas into something more checkable and concrete for agent-evaluation environments.

    There is also a gap between how complete the proposal initially feels and how much has actually been executed. R2 is still DESIGNED, R1 and R4 are PARTIAL, and the full end-to-end confined run and escape-attempt corpus are not yet complete.

    So I think this is a strong direction and a very good fit for the hackathon, but I would describe the current artifact more as a useful audit tool plus an emerging containment standard than as a fully validated standard.

    Read full reviewShow less

  2. This project brings attention to the practical containment gap of agents inheriting access that their own configuration don't reveal. The reported workstation audit helps prove this by including access from one agent's sandbox group to another agent's private directory. The evidence labels help separate completed checks from planned work. The permission audit is the clearest contribution at this stage, while the escape-detection harness and restricted-agent validation remain incomplete and the offered code was unavailable for inspection.

Cite this project

@misc{bauer2026containment,
  title = {{Containment at the Boundary: An OS-Level Standard and a Tested Escape-Detection Harness}},
  author = {Joshua Bauer},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/containment-at-the-boundary-an-oslevel-standard-and-a-tested-escapedetection-harness-8lph}},
  url = {https://apartresearch.com/sprints/projects/containment-at-the-boundary-an-oslevel-standard-and-a-tested-escapedetection-harness-8lph}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

AI x Epistemics Research Sprint · Nov 13 - 15, 2026