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

Watching the Boundary: Policy-Derived Detection and Containment for Autonomous Agent Evaluations

Dylan · Team Incident_Response

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

Read the report

Report: Watching the Boundary: Policy-Derived Detection and Containment for Autonomous Agent Evaluations

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

This project is an environment where an AI agent's actions are checked against a declared authorization policy. My main goal was for it to be deployable on both the OpenAI and Hugging Face sides of an incident like this. Hugging Face would use it somewhat like a honeypot, a world an attacker can be diverted into, where their traffic can be observed and analyzed safely. These incidents will keep happening, and a human may not be able to respond in time, even on the originating side. So OpenAI would use it to encapsulate their agent and get alerted the moment it tries to leave its authorized environment, while Hugging Face would use it for testing and for defense on their side.

Reviews

Judging this Sprint?

Review “Watching the Boundary: Policy-Derived Detection and Containment for Autonomous Agent Evaluations”

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 policy-derived authorization envelope is a compelling organizing idea: it turns expected agent behavior into enforceable boundaries, detection signals, and graduated containment actions. The project is strengthened by a substantive repository, reproducible run artifacts, multiple experimental conditions, prompt-injection testing, and a fast independent policy implementation.

    The evaluation would be considerably stronger with broader experimental coverage and clearer methodological detail. Please document the agent model, prompts, sampling settings, number of independent seeds, and how each reported result maps to a specific repository artifact. The prompt-injection result is encouraging, but nine trials across three variants are insufficient to establish robustness; additional injection families, indirect attacks, models, and repeated seeds would help. Likewise, the multi-agent experiment should cover more topologies, seeds, communication channels, and genuine value-transfer attempts.

    The latency result should be described more narrowly. The reported nanosecond measurement benchmarks the in-memory policy classification function, not the complete request path including parsing, networking, logging, alerting, and decoy routing. An end-to-end load test would better support production-readiness claims. Finally, clarify how the stated tamper-evident audit property is implemented, such as through hash chaining, signatures, or an external append-only store. These additions would turn a strong prototype into a much more convincing security evaluation.

    Read full reviewShow less

  2. The report is not very specific, It doesn't specify at all what the "series of controlled experiments" the framework was evaluated on entail, nor which LLM(s) was/were used in the experiments. It lacks any related work citation. Hard to judge the framework.

Cite this project

@misc{dylan2026watching,
  title = {{Watching the Boundary: Policy-Derived Detection and Containment for Autonomous Agent Evaluations}},
  author = {Dylan},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/watching-the-boundary-policyderived-detection-and-containment-for-autonomous-agent-evaluations-8gtr}},
  url = {https://apartresearch.com/sprints/projects/watching-the-boundary-policyderived-detection-and-containment-for-autonomous-agent-evaluations-8gtr}
}

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