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

Agent Worlds L3

SAHIL SONI

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

Read the report

Report: Agent Worlds L3

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

The July 2026 incident showed that autonomous agents can escape a sandbox through a permitted network channel to achieve their goals. I chose Track 1 (Containment), focusing on one rule: authority must only shrink during delegation. This rule prevents permission laundering, in which a low-privileged agent asks a more privileged agent to perform an unauthorized operation. I designed a draft conformance profile, a runnable reference harness, and an independent offline auditor. I tested five configurations: a shared-account baseline, a forwarded-credential sandbox, an Agent Worlds ablation, Agent Worlds, and scoped delegation. The shared-account baseline, forwarded-credential sandbox, and Agent Worlds ablation applied the laundered write, whereas Agent Worlds and scoped delegation blocked it. Attenuation is prior art; my contribution is independently checkable, signed evidence. The issuer signs with its private key, and the auditor uses the issuer's public key to detect invalid, incomplete, or contradictory supplied evidence. This mechanism does not prevent the initial network escape.

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. Good narrowly-scoped containment work. The rule is crisp (authority only shrinks during delegation) and the five-config comparison with an ablation is the right way to argue a conformance claim. Signed evidence plus an offline auditor is the best part, independently checkable audit trails are a real gap. But it is all simulated, one laundering scenario, so this shows the design is coherent, not that it survives a real sandbox. The trusted-recording assumptions need adversarial testing of their own. Next: one real deployment target and a second scenario.

  2. The project relies on a simulated python harness without real language models, network channels, containers, etc. This leaves various operational hurdles that could occur in the real world un-verified. More validation and a proper simulation based on real-world scenarios would make the paper more compelling.

Cite this project

@misc{soni2026agent,
  title = {{Agent Worlds L3}},
  author = {SAHIL SONI},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/agent-worlds-l3-bmmy}},
  url = {https://apartresearch.com/sprints/projects/agent-worlds-l3-bmmy}
}

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