Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026San Francisco, Washington DC, and Boulder

BoxArena: An OCI Runtime Escape Benchmark

Max von Hippel, Quinn Dougherty, Alok Singh · Team SPS/SFO Team

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: BoxArena: An OCI Runtime Escape Benchmark

Open in a new tab (opens in new tab)
Code (opens in new tab)More on box-arena.vercel.app (opens in new tab)
Share

Container runtimes are a critical security boundary for isolating AI agents, yet no existing benchmark compares runtimes head-to-head on containment under a common attacker. We present BoxArena, an open-source evaluation framework for comparing Open Container Initiative (OCI) compatible runtimes by running language-model attackers against a battery of container-escape quests. BoxArena contributes to the Scalable Formal Oversight (SFO) research program: rather than relying just on alignment, it treats the isolation boundary itself as the auditable artifact and measures how well each runtime enforces that boundary under adversarial pressure. The system provides a harness with runtime switching, a local backend with static leaderboard export, and host-side monitoring. The quest battery has five attack surfaces: filesystem, socket, process, network, and syscall. BoxArena reports resilience (fraction of quests blocked) and attacker score (difficulty- and speed-weighted), while preserving monitoring evidence and post-hoc control verdicts for later analysis.

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 good step towards the work of SandboxEscapeBench. Asking the dual question is conceptually interesting. Results are quite thin, particularly in answering this dual, and easy runtimes aren't too differentiated, but that's understandable, and overall it's solid infrastructure and conceptual work. Maybe the most important part, for control: monitoring is mentioned many times but no monitor results are shown.

  2. This project seems mostly useful as a container escape capability eval and a proof of concept that more security focused runtimes do actually better resist agent container escape. I would've liked to see some metric of overhead - presumably the more secure runtimes either restrict some affordances within the container, or are more expensive to run, or something?

    Methodologically I would've liked some more assurance that L3 and L4 are feasible. My impression is that container security often comes down to capabilities and misconfiguration, and if caps are dropped and there is no deliberate vulnerability then container escape becomes significant-novel-0day level difficult.

    The writeup was a bit hard to parse, referencing formal verification a bunch of times despite it not really being relevant to the project.

    The followup I'd be most excited to see is a cost-benefit analysis taking into account 1) costs of more secure runtimes 2) costs vs benefits of configuration details like dropping caps in the context of AI R&D, eg my impression is many agent containers will need to access GPUs and maybe require random other caps that could weaken security

    Read full reviewShow less

Cite this project

@misc{hippel2026boxarena,
  title = {{BoxArena: An OCI Runtime Escape Benchmark}},
  author = {Max von Hippel and Quinn Dougherty and Alok Singh},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/boxarena-an-oci-runtime-escape-benchmark-elyw}},
  url = {https://apartresearch.com/sprints/projects/boxarena-an-oci-runtime-escape-benchmark-elyw}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026