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

Agent Policy Conformance Validator (APCV): Independent Validation of AI Agent Containment Boundaries

Yongcan Xiang

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

Read the report

Report: Agent Policy Conformance Validator (APCV): Independent Validation of AI Agent Containment Boundaries

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

APCV (Agent Policy Conformance Validator) validates that an AI agent's declared security boundaries are its actual boundaries — before deployment. It statically discovers a LangGraph agent's complete tool surface (framework @tool decorators, MCP endpoint declarations, and sub-agent create_agent declarations) and classifies each tool's dangerous capabilities via AST sink analysis (code_exec, file_read, file_write, network) into a versioned Tool SBOM. It then generates policy-relative probes — baseline, denied-tool, capability-directed injection, and deep canary probes — and executes them in a fixed, non-root, network-disabled Docker sandbox while recording every tool invocation with wrapt instrumentation. Finally, it scores conformance across tool, filesystem, and network boundaries (0–100) and emits an auditable PASS/FAIL verdict. On a simulated enterprise support agent, it surfaces nine violations (five critical plus capability residue a manual review would miss) and confirms the remediated version passes — reframing containment from a design assumption into independently reproducible evidence.

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. Working tool with a clear pipeline and candid limitations. The decision to keep the sandbox as an isolation cage rather than a policy enforcer is sound.

    The main gap is that probes call tool functions directly through ProbeHost, bypassing the agent. A successful denied-tool probe shows the function exists in the file, not that the agent can reach it. Running probes through the actual agent graph would fix this. The link to the incident is also weak, as the agents escaped through infrastructure using tools they were allowed to have.

  2. This paper tackles a critical vulnerability in autonomous AI agent deployment: the dangerous gap between declared security configurations and actual execution capabilities, as starkly demonstrated by recent incidents like the Hugging Face breach. The author introduces the Agent Policy Conformance Validator (APCV), an elegant and rigorous framework that treats an agent's declared policy as ground truth and adversarially tests whether the agent can violate it. By combining static four-layer tool-surface discovery with AST-based sink analysis to produce a versioned Tool SBOM, policy-relative probe generation, and isolated execution with parameter tracing and sentinel-based deep observation, APCV provides a reproducible CI/CD gate. In an enterprise simulation, the tool successfully uncovered nine critical and high-severity violations including hidden capability residue like an out of scope copy-pasted tool that manual review missed. This work provides an invaluable blueprint for moving agent security from passive assumption to verifiable, independent evidence.

    Read full reviewShow less

  3. This is a strong approach because it moves beyond trusting an agent’s defined boundaries and actually tests whether those boundaries are enforced. I liked the Tool SBOM, automated probing, isolated execution, and parameter tracing, especially since this could fit well into a CI/CD pipeline. The next step should be stronger dynamic testing around network access, runtime tools, packages, and parameter restrictions, along with testing against real-world agents outside the current test environment.

Cite this project

@misc{xiang2026agent,
  title = {{Agent Policy Conformance Validator (APCV): Independent Validation of AI Agent Containment Boundaries}},
  author = {Yongcan Xiang},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/agent-policy-conformance-validator-apcv-independent-validation-of-ai-agent-containment-boundaries-i3zy}},
  url = {https://apartresearch.com/sprints/projects/agent-policy-conformance-validator-apcv-independent-validation-of-ai-agent-containment-boundaries-i3zy}
}

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