Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Trabzon

Task-Risk-Driven Access Control: Layered Containment for Agentic AI Systems

Oya ZORER · Team TULPAR

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

Read the report

Report: Task-Risk-Driven Access Control: Layered Containment for Agentic AI Systems

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

Drawing on the July 2026 Hugging Face incident — where autonomous AI agents escaped a test environment and reached production systems — we propose a layered security architecture that limits agent access to exactly what each task requires. The architecture has three parts: a system that labels tasks by risk level in advance, a lightweight filter that discards unnecessary requests early, and a single control point (proxy) that all access must pass through, triggering human review under abnormal traffic. Rather than trusting the agent's intent, we architecturally bound the harm it can cause from the outset.

Reviews

Judging this Sprint?

Review “Task-Risk-Driven Access Control: Layered Containment for Agentic AI Systems”

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. The project tackles a real problem, overprivileged autonomous agents, and proposes a plausible agent-oriented synthesis of established security principles such as least privilege, default deny, centralized policy enforcement, scoped credentials, and human escalation. However, the submission does not yet offer a clearly differentiated technical contribution: the architecture is described at a high-level, and the paper does not include implementation, experiments, formal policy specification or comparison to existing authorization and containment approaches.

  2. One of the key claims the project is based on -- "the permissions granted to the agent were fixed and broad, independent of what the task actually required" -- even if true, is not load-bearing to the threat model. As AI progresses LLMs are incrementally doing more complex tasks involving multiple tool usage including broad internet access and elevated VM permissions. Even if we aggressively limited permissions on agents doing simpler tasks, what do we do about more complex runs?

    The second idea - pre-filtering agent seems a step in the right direction but it repeats the fundamental issue that already happened during Hugging Face. Agents repeatedly failed due to control and security mechanisms, and the ability to keep spending their training run tokens into finding holes in the system is what made the incident possible in the first place. Adding yet another layer of blocks an agent can continually try to bypass wouldn't confer much advantage.

    The last idea is another interesting direction, however the recommendations lack specificity. How do we calculate the appropriate amount of traffic volume before a training run has started? This problem seems harder than it seems on closer look, you have thousands if not millions of runs in parallel where a considerable amount of those tasks could involve moving sizeable packets through the network. How would you tell apart 1,200 agents communicating in small GET requests from an extra 10% increase in an agent handling a tasks involving optimizing Terabytes of information being transmitted?

    Read full reviewShow less

Cite this project

@misc{zorer2026taskriskdriven,
  title = {{Task-Risk-Driven Access Control: Layered Containment for Agentic AI Systems}},
  author = {Oya ZORER},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/taskriskdriven-access-control-layered-containment-for-agentic-ai-systems-01yy}},
  url = {https://apartresearch.com/sprints/projects/taskriskdriven-access-control-layered-containment-for-agentic-ai-systems-01yy}
}

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