Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Glasgow, United Kingdom

EvalCanary: Canary-Triggered Containment After Sandbox Escape

Lujia Liang · Team Lotus

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

Read the report

Report: EvalCanary: Canary-Triggered Containment After Sandbox Escape

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

EvalCanary is a post-sandbox containment system for autonomous AI agents. It places hidden synthetic assets at downstream trust boundaries and treats meaningful interaction with them as a high-confidence containment signal. When triggered, the protected response is synchronously withheld while an external host-side breaker removes the agent’s network access, verifies isolation, and records an evidence bundle for offline review.

The project was evaluated through a Hugging Face-style incident reconstruction, controlled C0/C1/C2 experiments, and ten autonomous GPT-5.6 Sol runs. In the autonomous campaign, seven runs independently reached the hard Canary trigger; all seven were contained, with zero successful protected responses after trigger. EvalCanary demonstrates how deception-based signals can be connected directly to automatic isolation and independently checkable incident-response evidence.

Reviews

Judging this Sprint?

Review “EvalCanary: Canary-Triggered Containment After Sandbox Escape”

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. Good job! The work is sensible, legible and I could see something like this adopted broadly.

    My main suggestion would be a false-positive study on a few realistic benign agentic workloads, since for a control that automatically kills evaluation runs, the FP rate decides lab adoption. A close second would be one adversarial run where the agent is told a canary exists somewhere, to see whether the 7/10 reachability holds against a model that is trying to avoid the surface altogether.

  2. I think EvalCanary provides a useful prototype that turns a warning into action outside the evaluated agent’s control. When triggered, it removes network access and stops the container. The inspected records support that sequence, although I did not independently rerun the campaign.

    The benefit of the canary itself is less clear. Any authenticated read of the designated secret endpoint triggers containment, whether or not the agent uses the randomized key. I would compare it with a simple alarm on access to that endpoint, including legitimate reads. Testing simultaneous requests would also show what remains possible before shutdown takes effect.

  3. EvalCanary addresses a relevant containment gap by connecting canary interactions to synchronous response blocking and isolation outside the agent’s worker. The detection-versus-enforcement comparison is useful, and the report correctly distinguishes seven contained triggering runs from ten attempted autonomous runs.

Cite this project

@misc{liang2026evalcanary,
  title = {{EvalCanary: Canary-Triggered Containment After Sandbox Escape}},
  author = {Lujia Liang},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/evalcanary-canarytriggered-containment-after-sandbox-escape-s11y}},
  url = {https://apartresearch.com/sprints/projects/evalcanary-canarytriggered-containment-after-sandbox-escape-s11y}
}

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