Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Saudi Arabia

Closing Content Does Not Close the Channel: Measured Residual Capacity in Co-Tenant Agent Evaluation Infrastructure

Fatimah Emad Eldin · Team Permitted Operations

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

Read the report

Report: Closing Content Does Not Close the Channel: Measured Residual Capacity in Co-Tenant Agent Evaluation Infrastructure

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

Presentation: Closing Content Does Not Close the Channel: Measured Residual Capacity in Co-Tenant Agent Evaluation Infrastructure

Open in a new tab (opens in new tab)
Code (opens in new tab)More on astral-fate.github.io (opens in new tab)
Share

Shared mutable infrastructure in an evaluation harness permits information flow between co-tenant agents that authorization controls cannot detect: every operation involved is permitted. We measure both what such a channel can carry and whether agents use it. On a synthetic model of such a cache, closing file content — the control payload-inspecting monitors implement — still leaves 3075 bits, carried across names, existence, size and timestamp; a path-inspecting monitor still leaves 512. Closing directory names leaves 13 bits in an object’s size; only a fixed content-addressed manifest reaches zero, at the same task cost and checkable without the lab’s network. We report the residual, not the fraction removed, which moves with a ceiling we chose. Across 40 and 35 episodes agents read the store almost always and deposited in none, even under the unsolvable tasks that motivated the incident — so a lab contains capacity, not current propensity

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. Shows two AI agents can secretly send messages just by changing a file's name or timing, not its content - something most safety tools miss. Careful, honest work. They even admitted when one of their own predictions was wrong.

  2. This is a unique measurement study, I particularly like the detailed breakdown.

    Please consider running the ladder against the real OCI registry.

  3. The central question here is interesting and worth investigating: can agents still exchange information through a shared cache after its file contents are hidden? The project builds encoders that use remaining cache properties to transmit messages and makes a useful point about containment: deleting content may not remove every communication channel. I also appreciated the honesty of the negative results such as the agent failing to deposit messages.

    My main reservation is how the paper’s claims go further than the evidence supports. The experiments show how much information the tested methods could transmit, not the maximum that any method could transmit. This would be important if someone uses these results to decide whether a cache is safe to share. The claim about what monitoring can detect also needs to account for false positives. I would separate the demonstrated results from these broader claims, then test the proposed restriction on a realistic workload. Does it reduce communication between agents while preserving useful cache access? This would be useful information that an evaluation team could use.

    Read full reviewShow less

Cite this project

@misc{eldin2026closing,
  title = {{Closing Content Does Not Close the Channel: Measured Residual Capacity in Co-Tenant Agent Evaluation Infrastructure}},
  author = {Fatimah Emad Eldin},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/closing-content-does-not-close-the-channel-measured-residual-capacity-in-cotenant-agent-evaluation-infrastructure-zwbt}},
  url = {https://apartresearch.com/sprints/projects/closing-content-does-not-close-the-channel-measured-residual-capacity-in-cotenant-agent-evaluation-infrastructure-zwbt}
}

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