Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 12, 2026Waterloo, Ontario, Canada

SEP-1: A Sole-Egress Package-Broker Profile for Model-Evaluation Sandboxes

Zhixuan Xu (George) · Team UWaterloo

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

Read the report

Report: SEP-1: A Sole-Egress Package-Broker Profile for Model-Evaluation Sandboxes

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

In July 2026, the OpenAI→Hugging Face escape left the sandbox through the only network exit it was allowed to have: the package-registry proxy. The egress policy never even fired, because it was watching destinations, not payloads. Where the sandbox connected looked fine. What actually crossed the boundary didn't matter to it.

SEP-1 is our attempt at a containment standard that drops the allowed-destination model altogether. Instead of a network, the sandbox gets a socket. A broker on the other side serves only pre-resolved, hash-pinned packages, and the sandbox itself has no network interface at all — nothing to route, nothing to escape through.

To test this, we built a harness and ran the same 12-attack suite twice, once against an allowlist setup and once against SEP-1, with a compromised upstream in both cases. The results were lopsided. Under the allowlist, 11 of the 12 attacks made it across the boundary — and the policy monitor still called 13 of 14 connections "in-policy." SEP-1 contained 11 of 12. The twelfth degrades into a covert channel, which we measured and bounded: 216,000 bits/hour worst case, and roughly 15 bits per sandbox lifetime once caching kicks in.

Each run ends in a hash-chained conformance report. A third party can verify it against the public lockfile and probe suite without touching the lab's network. An allowlist log can't give you that, and we think this auditability — not the attack numbers alone — is what makes the standard something a regulator could actually check and adopters could actually trust.

Reviews

Judging this Sprint?

Review “SEP-1: A Sole-Egress Package-Broker Profile for Model-Evaluation Sandboxes”

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. The project presents a concrete package-broker design that restricts retrieval to approved, hash-pinned artifacts. Its mock-upstream comparison reports containment of 11 of 12 probes while preserving legitimate retrieval and exposing a residual communication channel. The design has practical potential, but its leakage bounds and deployment guarantees need further validation.

    Strength: Restricting the sandbox to approved artifacts addresses a concrete weakness in destination-based controls. The successful legitimate-fetch test makes the design more useful by showing that containment can preserve an intended function.

    Recommendation:

    • Define the security promise precisely. Blocking arbitrary downloads and preventing sensitive information from leaving are different guarantees. The remaining request channel makes that distinction consequential: the caching estimate considers ordering, while request timing may reveal additional information. Define the complete observable channel and explain what level of leakage is acceptable for the intended evaluation. This would turn the reported bound into a meaningful security assurance.
    • Demonstrate the guarantee in the deployment being proposed. The committed results use in-process orchestration, while the design relies on enforced network isolation. An end-to-end container demonstration that preserves legitimate retrieval while denying other network access would provide stronger evidence that the profile works as an operational containment control.
    Read full reviewShow less

  2. This project effectively eliminates egress escapes by removing virtual network interfaces entirely and relying solely on a Unix Domain Socket (UDS) governed by SEP-1's eight validation rules (pinned SHA hashes, size checks, and redirect blocking). It is an elegant architectural shift, though I wonder how cleanly this will scale across high-concurrency workloads and distributed multi-node clusters. That said, trading a slight latency hit at startup for deterministic AI containment is an entirely reasonable compromise

  3. SEP-1 addresses a practical problem because evaluation agents may need software packages without needing unrestricted internet access. It proposes choosing approved package versions in advance and making them available through a restricted service. That could reduce opportunities for agents to communicate outside the evaluation while preserving useful work.

    The claimed limits on communication need stronger support. An agent could communicate through when it requests packages, not just which packages it chooses. The report also needs to make clearer which protections were implemented and which remain proposed. I would compare a realistic package dependent task under SEP-1 with the same task using preinstalled packages and no network access. That would help establish whether the extra service offers enough benefit to justify its complexity.

Cite this project

@misc{george2026sep1,
  title = {{SEP-1: A Sole-Egress Package-Broker Profile for Model-Evaluation Sandboxes}},
  author = {Zhixuan Xu (George)},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sep1-a-soleegress-packagebroker-profile-for-modelevaluation-sandboxes-7kif}},
  url = {https://apartresearch.com/sprints/projects/sep1-a-soleegress-packagebroker-profile-for-modelevaluation-sandboxes-7kif}
}

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