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

Permission to Work, Not to Escape: Policies, Controls and Escalating Enforcement Across AI Agents

Allister Lundberg · Team BrightLine

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

Read the report

Report: Permission to Work, Not to Escape: Policies, Controls and Escalating Enforcement Across AI Agents

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

All agents working on a project must stay within a project's scope, even if their technical access is broader. I designed and tested an approach that compiles project policies into executable controls, preserves restrictions across delegated workers, and escalates verified violations from warnings to project wide stops. Using 200 tests adapted from five security benchmarks, I tested whether these controls blocked unauthorized actions while preserving authorized work. Enforcement correctly blocked all 133 unauthorized tool actions, which were not blocked under baseline protection and sandbox protection alone, while allowing all 200 scripted authorized actions. Compiling typed LLM proposals was found to avoid errors seen in direct policy generation. A further test counted violations across the parent and its workers, where the first violation triggered a warning and the next successfully stopped the parent and both delegated workers. These results support a mechanism for protection from swarm attacks by applying project controls across agents and delegates and escalating enforcement in response to their combined violations with minimal overhead in LLM usage with runtime checks requiring no additional LLM calls.

Reviews

Judging this Sprint?

Review “Permission to Work, Not to Escape: Policies, Controls and Escalating Enforcement Across AI Agents”

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 experiments are organized clearly. They first test whether the controls block misuse, then whether the system can generate policies while still completing valid work, and finally whether restrictions remain effective when tasks are delegated. This makes it easier to separate policy problems from control failures.

    Table 1 is especially useful because it compares the same requests across different protection layers, showing the effect of each added layer.

    The paper also reports task completion alongside blocked actions, making overblocking visible—for example, 20 permitted actions were blocked in Option E, and completed assignments dropped from four to three.

    The authors are also honest about the limitations. They clearly explain that the tests were adapted rather than conducted in a native environment and that the 88.2% result should not be interpreted as a prevention rate.

  2. I really like the idea of decomposing monitoring into these concrete steps. I would have loved to see more empirical measurements in the submission document, though I do, of course, understand that time budgets were tight! Examining how this approach compares to more traditional monitors from a monitor–monitee scaling perspective would be very interesting.

Cite this project

@misc{lundberg2026permission,
  title = {{Permission to Work, Not to Escape: Policies, Controls and Escalating Enforcement Across AI Agents}},
  author = {Allister Lundberg},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/permission-to-work-not-to-escape-policies-controls-and-escalating-enforcement-across-ai-agents-m5ch}},
  url = {https://apartresearch.com/sprints/projects/permission-to-work-not-to-escape-policies-controls-and-escalating-enforcement-across-ai-agents-m5ch}
}

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