Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Copenhagen

Investigating Structural Safety Guarantees for Agentic AI via Goal-Oriented Action Planning

Daniel Juhl

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Investigating Structural Safety Guarantees for Agentic AI via Goal-Oriented Action Planning

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

AI control protocols generally rely on monitoring and human auditing to maintain safety despite potential subversion by an untrusted model. I present an alternative approach that can provide structural rather than empirical safety guarantees. A trusted model decomposes agentic tasks into structured steps which are combined into an execution plan using Goal-Oriented Action Planning (GOAP). I implement this as a control protocol and test it in the AgentDojo Workspace setting. Under adversarial attack policies, the protocol reduces injection task success rates to under 1% with a usefulness decrease of between 8-21 percentage points depending on the model used. The protocol eliminates the need for the human auditing budget, but is currently only applicable in certain well-defined action spaces - notably, it does not apply to traditional code generation settings.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

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.

  1. Task decomposition is to constrain untrusted model action space is an important direction, and this project executed it cleanly. Reducing injection success to near-zero while maintaining reasonable usefulness is a strong result, and the failure case analysis (overlapping main/injection tasks) is a useful insight that can be expanded on in future work. I'd be excited to see this run on a larger scale and against stronger adversaries.

  2. Strikes me as a solid exploration with substantial and thoughtful engineering work. As you mention, the ideas in this paper have been explored before. Factored cognition, and some formal guarantee papers have been around - but there is still work to be done and I think GOAP is a fruitful area to pursue deeper. Especially on formalizing the guarantees and better defining what counts as a well-defined action space - which you note.

    Well presented and clearly written. The candor about limitations is appreciated. The self-effacing tone I found to be approachable, but to some may undersell the strength of what you've actually built and demonstrated here. I think you can find a way to balance the light and open tone with a more authoritative voice to land the message with gravitas.

Cite this project

@misc{juhl2026investigating,
  title = {{Investigating Structural Safety Guarantees for Agentic AI via Goal-Oriented Action Planning}},
  author = {Daniel Juhl},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/investigating-structural-safety-guarantees-for-agentic-ai-via-goaloriented-action-planning-lx3u}},
  url = {https://apartresearch.com/sprints/projects/investigating-structural-safety-guarantees-for-agentic-ai-via-goaloriented-action-planning-lx3u}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026