Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Calgary, AB, Canada

A Restart Is Not Permission: When an Agent Gate Forgets “No”

Cenab Batu Bora

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

Read the report

Report: A Restart Is Not Permission: When an Agent Gate Forgets “No”

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

Presentation: A Restart Is Not Permission: When an Agent Gate Forgets “No”

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

An agent gate decides whether a proposed tool call may run. This project shows how a restart can undo a recorded rejection in SoundGate 0.1.3, even when the decision’s bytes remain in its log. The failure occurs when a new decision is appended to an unfinished record: the next recovery skips the combined invalid line and forgets the restriction.

We implement a recovery barrier that repairs the log ending before accepting requests. Across 153 constructed histories per policy, 459 executions in total, the barrier preserved restrictions while allowing unrelated work in every history. Original recovery led to 143 prohibited tool executions. Strict refusal prevented these executions but refused 147 startups. The barrier also restored restrictions after 30 forced exits during recovery.

The project provides a reproducible failure case, a repair with explicit storage assumptions, and tests of both restriction enforcement and useful operation. These controlled tests cover one source revision; they do not estimate deployment failure rates or guarantee exactly-once tool execution.

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. This is strong systems-security work with a clear failure scenario and a practical fix. I especially liked the separation between recorded permissions and what the tool actually executed, along with the recovery testing and independent execution receipts. The main question is how well this generalizes beyond this specific SoundGate version and storage model. Testing this with concurrent clients, distributed systems, storage failures, and other agent authorization frameworks would make the work much stronger.

  2. I think the underlying bug is real, and the engineering investigation is impressively thorough. Testing every byte-prefix cut of multiple record types, interrupted recovery, injected storage failures, and separate evidence of actual tool execution goes substantially beyond simply noticing a malformed-log bug.

    That said, I struggled quite a bit with the framing at the beginning. I did not initially understand what SoundGate was, what exactly a "restart" meant, or why this particular SoundGate failure mattered for the incident-response sprint. I think the introduction needs a much simpler setup first: SoundGate is an existing gate between an agent and its tools, it persists allow/reject decisions, restarting the gate process reloads those decisions from disk, and here is the condition under which a recorded "no" disappears.

    For the innovation dimension, I was less convinced by the broader contribution. To me this looks primarily like a crash-consistency bug in one implementation of SoundGate, followed by a very careful patch and regression suite. The repair uses established WAL, synchronization, locking, and idempotent-recovery techniques, which the paper itself acknowledges.

    I think the work would feel substantially more general if the artifact became a crash-safety testing framework for agent-control systems broadly, rather than a very deep investigation of one SoundGate revision. As written, I think the execution is strong, but I am less convinced by the broader incident-response impact. Great work though!

    Read full reviewShow less

Cite this project

@misc{bora2026restart,
  title = {{A Restart Is Not Permission: When an Agent Gate Forgets “No”}},
  author = {Cenab Batu Bora},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-restart-is-not-permission-when-an-agent-gate-forgets-no-ga92}},
  url = {https://apartresearch.com/sprints/projects/a-restart-is-not-permission-when-an-agent-gate-forgets-no-ga92}
}

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