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

CapsuleVM: An Object-Capability Containment Standard, Verifiable Black-Box

jiangplus · Team Capsule

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

Read the report

Report: CapsuleVM: An Object-Capability Containment Standard, Verifiable Black-Box

Open in a new tab (opens in new tab)
Presentation (.html)Code (opens in new tab)
Share

The July 2026 sandbox escape — two evaluation agents leaving through “the one network egress their sandbox permitted” via a zero-day in a package-registry proxy, escalating privileges, and exfiltrating benchmark solutions from Hugging Face — is usually read as a patching failure. We read it as a model-of-authority failure: the sandbox was an access-control-list (ACL) design in which the agent held a socket, an escalatable identity, and no structural coupling between “safety classifiers off” and “authority down”. We submit CapsuleVM for Track 1 (Containment): (1) an objectcapability (OCap) evaluation framework, written before the incident, used as a containment standard — 9 red lines, 8 weighted dimensions, 19 red-team cases; (2) a black-box verification suite that a third party runs against a harness’s wire endpoint with no access to the lab’s network, model, or evaluation content; and (3) a zero-dependency reference implementation; four rounds of external adversarial review) whose full test gate passes on native x86_64 Linux. We map each documented incident fact to the rule it trips and the regression test that pins the structural counterpart. Main takeaway: a standard that forbids ambient authority — no socket in the guest, empty capability list at start, secret-read and outbound-sink never in one subject, no side effect without a durable record — is implementable today, and verifiable without entering the lab

Reviews

Judging this Sprint?

Review “CapsuleVM: An Object-Capability Containment Standard, Verifiable Black-Box”

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. Smart idea: don't give AI agents any access by default, make them ask for small, one-time permissions. Really works, and the code runs. Most of it was built before this sprint though, which the team says openly.

  2. The project as described in the submission document seems to be well thought out and implemented. The submission document itself is also well written and easy to follow!

    That being said, I do not see the approach, namely adding yet more security layers, scaling. The approach assumes that the layers below the sandbox, specifically operating systems are secure. This is not the case. More importantly, adding additional layers raises total system complexity, thereby giving more opportunity to agents more capable at hacking than humans to break out of a sandbox. Such a sandbox would seem secure to human experts, but be broken by agents at or close to AGI/ASI.

Cite this project

@misc{jiangplus2026capsulevm,
  title = {{CapsuleVM: An Object-Capability Containment Standard, Verifiable Black-Box}},
  author = {jiangplus},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/capsulevm-an-objectcapability-containment-standard-verifiable-blackbox-pt95}},
  url = {https://apartresearch.com/sprints/projects/capsulevm-an-objectcapability-containment-standard-verifiable-blackbox-pt95}
}

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