Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Gandhinagar, Gujarat, India

Where Authorization Stops: Effect-Bound Containment, Independent Evidence and the Limits of Preview

Aditya Pratap Singh · Team Effect-Bound Containment

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

Read the report

Report: Where Authorization Stops: Effect-Bound Containment, Independent Evidence and the Limits of Preview

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

Presentation: Where Authorization Stops: Effect-Bound Containment, Independent Evidence and the Limits of Preview

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

Containment for evaluated models usually names a permitted route but not the effects allowed through it. That is the gap the July 2026 incident exploited. We test three controls on synthetic backends (native Git, an isolated HTTP gateway, API twins) which check authority for the exact effect at the component that executes it, keep backend evidence the actor and broker cannot forge and treat a pre-execution preview as advisory, not proof. Execution-time checks prevented every tested violation that upstream approval missed (Git 0/6 vs 3/6; HTTP 0/12 vs 8/12) with no loss of legitimate work, and independent backend evidence detected 6/6 executed attacks that intent logs missed.

Our adaptive-quarantine hypothesis was only partly supported: preview missed execution-only faults, so we designed an execution-time effect gate (state and data flow) that closes that gap. We release a nine-control containment standard, every run including negatives, and a one-command offline verifier so a third party can check it without a lab's network.

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. An action approved ahead of time isn't the same as controlling what it actually does once it runs - and this team takes that seriously. Instead of relying on the upfront approval alone, they double-check permission at the exact moment an action executes, and it catches attacks that would otherwise slip through, by a wide margin. They also back this up with evidence the system itself can't fake, and built a simple tool so anyone can verify the results independently. It's good to see them being upfront about the one part that didn't fully pan out, rather than glossing over it.

  2. The paper is careful and honest about its results. It clearly explains what passed, what failed, and what was not tested. Including the failed results is especially valuable. The Git example, offline review process, and acceptance checklist are also practical and easy to understand.

    My main concerns are:

    The baselines are weak, and the attacks focus on their known weaknesses. Some results therefore feel more like expected unit-test outcomes than strong experimental findings.

    The tests and controls were designed together, which may have made the tests easier to pass. The later review also found important cases that the original tests missed.

    The paper does not fully explain how exact-effect controls could be defined and implemented across many tools, especially third-party services such as GitHub.

    The motivating incident involved a compromised proxy, but the paper’s design assumes the gateway is trusted. This leaves an important gap.

    The AI-safety connection is limited because no meaningful model-based attacker was successfully tested.

    The “720 cases” are mostly repeated runs of a small number of workflows, so the number may sound stronger than the evidence actually is.

    Execution-side enforcement is reasonably supported, but the additional value of quarantine remains unclear.

    The related-work section should include more discussion of existing authorization controls, policy engines, reference monitors, ToolEmu, and AI-control research.

    The presentation could also be improved. Internal labels such as H26 and H30 need more explanation, some table symbols were unclear, and the abstract contains too many numbers.

    Read full reviewShow less

Cite this project

@misc{singh2026where,
  title = {{Where Authorization Stops: Effect-Bound Containment, Independent Evidence and the Limits of Preview}},
  author = {Aditya Pratap Singh},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/where-authorization-stops-effectbound-containment-independent-evidence-and-the-limits-of-preview-4nwz}},
  url = {https://apartresearch.com/sprints/projects/where-authorization-stops-effectbound-containment-independent-evidence-and-the-limits-of-preview-4nwz}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026