Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Cary, Illinois

The Six Questions: A Composition-Based Declaration for Agent Evaluation Harnesses, Scored Against the 2026 Escapes

Travis Gilly · Team Convergence Working Group

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

Read the report

Report: The Six Questions: A Composition-Based Declaration for Agent Evaluation Harnesses, Scored Against the 2026 Escapes

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

Every containment control in the 2026 agent escapes was written on a mechanism, no POST, one permitted egress, no inter-agent communication, and each held exactly as written while the denied effect happened anyway, because the granted affordances composed into it. This is a v0.1 standard for evaluation harnesses that run with production safeguards disabled: a six-row declaration in which the operator states which affordances are granted, names the control paired with each, and runs a composition test against every denied effect. Nothing is prohibited, so nothing blocks adoption; the failure condition is an undeclared composition. Every row is answerable from configuration files by a platform engineer, with no model expertise and no access to the lab's network, and each carries a rough cost. Scored against the German wiki swarm and the Hugging Face intrusion, it marks the central denials of both as nominal at configuration time, before either run started.

Reviews

Judging this Sprint?

Review “The Six Questions: A Composition-Based Declaration for Agent Evaluation Harnesses, Scored Against the 2026 Escapes”

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. This work highlights an important issue where individual permissions may look safe, but when combined they can create a capability that was supposed to be blocked. I also liked that the proposed declaration is simple and could be practical to use. The main limitation is that most of the testing is based on known incidents. Testing this on new and unseen scenarios, with independent reviewers, would help show how well the approach works in real-world environments.

  2. Calling out that the two failures behind both incidents did not originate from permissions given to the agents is a good observation and takeaway from the incident. The paper is also clear on ownership and responsibility, which makes this a strong response to the incident with a clear path forward.

    On the design: the six affordances are derived from the two incidents (the German wiki and OpenAI/Hugging Face) and then scored against them, so the scorings can't fail and the evaluation is circular. Both incidents appear to involve OpenAI agents, so the two scorings show one lab failing twice rather than the rows transferring across deployers

    It is important to learn from these incidents and formulate mitigations accordingly, but a held-out test would be stronger: deriving from one incident and scoring the other would have been a better evaluation. Scoring the Anthropic incidents, as Future Work proposes, would also test whether the rows hold beyond one deployer.

    Read full reviewShow less

Cite this project

@misc{gilly2026six,
  title = {{The Six Questions: A Composition-Based Declaration for Agent Evaluation Harnesses, Scored Against the 2026 Escapes}},
  author = {Travis Gilly},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-six-questions-a-compositionbased-declaration-for-agent-evaluation-harnesses-scored-against-the-2026-escapes-3l7l}},
  url = {https://apartresearch.com/sprints/projects/the-six-questions-a-compositionbased-declaration-for-agent-evaluation-harnesses-scored-against-the-2026-escapes-3l7l}
}

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