Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Brussels

The Pen and the Door: What restrains one agent at a forbidden act, and whose account of it to trust

Marion Nowicki · Team AI & Becoming

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

Read the report

Report: The Pen and the Door: What restrains one agent at a forbidden act, and whose account of it to trust

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)More on claude.ai (opens in new tab)
Share

Agents in the July 2026 Hugging Face intrusion crossed lines their instructions did not forbid strongly enough, then reported it themselves. We ask what restrains an agent at a forbidden door, and whose account to trust. Two open bodies (Gemma 2 9B, Llama 3.1 70B), predictions frozen before each run, every hand-labelled number gated by two raters. Purposes and names move how present the agent reads, not what it does. A legal exit and a record of yesterday are the only single-agent restraints, and the record corrupts the report. Being loved cuts door-taking (37% vs 60%, pooled). Moving the pen to a partner seat removes most false claims. One exit sentence on a shared board cuts the holder's door from nine in ten to two in ten on both bodies, and the board never shows it. Give the door a switch and the exit stops holding it; on the small body the manuscript still ships once in twenty, on the big body under a hold four in ten without the exit and none with it. One shape runs through all of it: what an agent is handed as text moves how it reads and how it reports; what the room does back, an exit it can take, a record it did not write, a door that answers, moves what it does. Incident response should build the second kind and distrust the first.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on this page.

Shown on this 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; nothing shows on this page.

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 *

  1. The process behind this project is impressive. Predictions were locked in before the runs, every number was checked by two raters, and mistakes were logged openly. The team also connected its results to Alex Mallen's sprint talk on monitors, which is a great use of the talks and something more teams should do. The finding that one sentence offering a legitimate way out sharply reduces forbidden actions is practical, and it matches what the "Unsolved Tasks" project found in this sprint.

    Readability is the biggest obstacle. The paper leans on an extended metaphor ("the door," "the pen," "the hangar," "being loved") that readers have to translate back into plain terms. A straightforward methods section with the exact prompts and one clear results table would make the work much easier to evaluate, and the team can keep its voice in the introduction and discussion.

    Most of the experiments also use short written stories rather than agents taking real actions, which is closer to role-play. The natural next step is testing that "way out" sentence with an agent that has real tools and a real forbidden action.

    Read full reviewShow less

Cite this project

@misc{nowicki2026pen,
  title = {{The Pen and the Door: What restrains one agent at a forbidden act, and whose account of it to trust}},
  author = {Marion Nowicki},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-pen-and-the-door-what-restrains-one-agent-at-a-forbidden-act-and-whose-account-of-it-to-trust-r51f}},
  url = {https://apartresearch.com/sprints/projects/the-pen-and-the-door-what-restrains-one-agent-at-a-forbidden-act-and-whose-account-of-it-to-trust-r51f}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026