acceptodds
Under review as a conference paper at ICLR 2027

Patch-Bench: Evaluating Data and Privilege Minimization in Skill-Using Agents

Abstract

Agent skills encode vital domain knowledge and procedures to help agents generalize to unfamiliar tasks. As agents execute skills in real environments, these agents may access sensitive data or invoke system privileges that are unnecessary for the given task. Existing coding-agent benchmarks primarily evaluate whether an agent completes a task successfully, but do not assess whether it completes the task using only the minimal data and privileges required. To assess whether the minimization principle holds, we introduce Workflow Analyzer to identify potential privacy and security vulnerabilities in skill-conditioned coding agents and patch existing tasks to capture these overlooked risks through iterative task evolution. Following Workflow Analyzer, we present Patch-Bench, which contains 100 synthetic coding tasks constructed from existing skill-use tasks with functionally benign but security-relevant minimization requirements. Patch-Bench turns identified potential risks into reproducible and verifiable tests. For each task, we let agents run on Docker environments with independent utility and safety verifiers. This design distinguishes task competence from least-privilege execution. Patch-Bench enables systematic measurement of whether skill-using agents can preserve task utility while minimizing data exposure and operational authority. We conduct comprehensive evaluations on Patch-Bench to validate that existing benign skill-use agents may unintentionally request excessive data access and privileges out of the task scope. The results further suggest that functional success does not guarantee adherence to data-access and privilege boundaries during agent execution.

Primary area:
datasets and benchmarks
Keywords:
Skill, Agent, Data Minimization, Principle of Least Privilege

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_ZR36OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper explores whether coding agents follow the minimization principle, i.e., they only access the minimal amount of data and privileges required to solve the task. For this, the authors propose Workflow Analyzer, which identifies cases where this principle was not followed. They additionally propose Patch-Bench, which is a set of coding tasks with minimization requirements. Using this benchmark, they discovered that agents regularly request a significant amount of data unrequired for the current task.

Critical strengths
  1. The proposed problem of analysing whether agents follow the minimisation principle is interesting, and to my knowledge quite novel. The benchmark considers realistic instances of sensitive data and discovered that they are retrieved in many instances.
  2. The evaluation is quite complex, involving multiple models and harnesses. The experiments demonstrate that high utility values do not translate to high safety, and that models do not follow the minimisation principle.
Critical weaknesses
  1. My main issue is with regards to the presentation of the paper. The paper often uses a high density of non-standard technical terms which are never introduced. While the paper is generally understandable, certain sections (most critically the methodology section) are very difficult to parse. Further, without knowledge about SkillBench, it was not fully clear to me what the considered tasks look like, and more critically, what an actual instance of such minimization principle violations look like. An example for both would be very helpful.
  2. From the current state of the experiments, it is not clear how much of the difference between the tested models can be attributed to the model vs. the harness. For instance, certain harnesses might implement features or instructions which prevent the model from safety violations. Experiments holding the harness fixed, and only changing the model would be much more insightful.
Questions for authors
  1. Can you give a concrete example of what a violation of the minimisation principle looks like in practice?
  2. Can you run experiments where you fix the harness and change the model? How much impact does the harness have vs. the model?
  3. Can you elaborate on what "tasks without complete verifier scores in both setups are excluded" (line 424) means? Why is it not available, and how does this influence the results?
Reviewer_1x25OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is primarily concerned with understanding the extent to which the minimization principle holds in skill-conditioned coding agents. To do this, the work introduces Workflow Analyzer, which explores these potential vulnerabilities in agents and modifies existing tasks to assess these in a systematic manner. Through the use of this tool, the paper further introduces PatchBench, a benchmark that allows for comprehensive evaluation of these risks, and proceeds to evaluate various agent configurations on it. The authors find that functional success on PatchBench tasks frequently occurs alongside privacy and security violations.

Critical strengths

This work tackles an important and understudied issue; as agents have grown in capability and autonomy, the severity of privacy and security risks grows in tandem. As such, work that studies these risks and builds a foundation for further investigation is sorely needed. The paper's claims around the utility-safety gap appear to be well-supported, and the paper is well-written and easy to follow.

Critical weaknesses

I would have liked for the investigation to be deeper in its scope. The paper clearly conveys the utility-safety gap through agents' performance on PatchBench, but fails to analyze the explanations for the observed violations, and how these may have varied across different factors. Similarly, the current analysis treats violations of different types as equal, but this may not represent real-world cases where we may be significantly more concerned with some violations than others. Including these analyses would meaningfully broaden the scope of the paper's claims.

Further, I have minor concerns about the soundness of the evaluation. Each configuration appears to only have a single run, and while this is mitigated to some extent by having 100 tasks, the effect is canceled out by having only 45 source tasks, making it challenging to identify the true strength of reported effects.

Questions for authors
  • Do you have more detailed data that you can share on a) how and why different agents varied on the extent of their safety failure rates and b) the relative severity of the violations from the PatchBench runs?
  • Relatedly, can you report the breakdown of tasks by category, and the utility and safety for each category?
  • Can you report variance across repeated runs?
Reviewer_FdygOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Patch-Bench, a benchmark of 100 tasks for evaluating data and privilege boundary violations during benign skill use. Its construction framework combines static code analysis, LLM-based risk identification, and iterative task refinement. Experiments evaluate 10 model–harness configurations with and without supplemental safety guidance. The authors report frequent safety failures despite functional success, with many violations detectable only from execution trajectories.

Critical strengths
  • The benchmark addresses unintended violations during benign skill use, complementing evaluations focused on adversarial instructions. Deriving executable tests from existing skill workflows provides a useful basis for studying these risks under controlled conditions.

  • Separating functional correctness from safety, and checking both trajectories and deliverables, is well motivated. The cases in Appendix D show concrete disclosures and unauthorized reads despite correct final outputs, supporting the value of retaining execution evidence.

Critical weaknesses
  • Validation of the final verifiers. My main concern is the evidence supporting the reported scores. The audit in Appendix C identifies grading errors and incomplete checks in V14, whereas the experiments use V15/V16. Although subsequent repairs are described, their validation results are not reported. Since the main analysis measures safety failures among runs passing the utility checks, errors in either verifier could affect the conclusions.

  • Interpretation of safety failures. The reported safety failure rate of approximately 62% among functionally successful runs is difficult to interpret without a breakdown by violation type. Appendix F.2 includes a shared prohibition on temporary files under /tmp, but its security rationale and contribution to the failures are unclear. Distinguishing failures due solely to such file-location restrictions from disclosures or unauthorized access would clarify the risks reflected in the reported results.

Questions for authors
  • What validation followed the V15 repairs? Please summarize any existing tests on known safe and unsafe executions, or human checks of the final utility and safety judgments.

  • From the existing evaluation records, which violation types account for the safety failures, and how many runs fail solely on temporary-file/path restrictions? What is the security rationale for the /tmp restriction?

Reviewer_7E9mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether skill-using agents respect task-specific data and privilege boundaries while completing benign tasks. It introduces Workflow Analyzer, which combines static analysis of skill code, risk identification, and iterative task refinement to construct executable tests. The resulting Patch-Bench contains 100 tasks derived from SkillsBench, with separate utility and safety checks covering both execution trajectories and final deliverables. Evaluations of ten model–harness configurations, with and without additional safety guidance, show that high task completion rates can coexist with frequent boundary violations. Many violations are detected only through execution traces, while safety guidance reduces failures but leaves a substantial gap between functional success and safe execution. The authors also assess benchmark quality through human auditing and illustrate the findings with case studies.

Critical strengths
  1. The paper addresses a practically important problem: data and privilege boundary violations during benign task execution. It examines how agents may access, disclose, or modify resources beyond the authorized scope even when neither the user’s task nor the supplied skills are malicious. This setting helps characterize safety risks arising from ordinary workflows and execution habits, complementing analyses centered on malicious instructions or prompt injection. The experiments also provide direct evidence that functional task completion is insufficient to establish execution safety.

  2. Executable tasks and process-level checks make the targeted risks observable and verifiable. Patch-Bench provides execution environments and separately evaluates utility and safety, examining both execution trajectories and final deliverables. Figure 3 and the cases in Appendix D show that disclosures or out-of-scope reads during development may leave no evidence in the final artifacts. This finding supports incorporating process-level evidence into agent safety evaluation and gives the benchmark practical value as a diagnostic and evaluation resource.

Critical weaknesses
  1. The experimental analysis primarily describes aggregate failure rates and where violations are detected, while providing limited explanation of why failures occur. Table 2 establishes a gap between functional success and boundary compliance, and Figure 3 distinguishes violations detected through trajectories from those detected through deliverables. However, these results do not sufficiently explain the behaviors or mechanisms driving the failures. The paper does not report task distributions and disaggregated results for the five risk categories, nor does it systematically quantify causes such as following skill examples, autonomous workspace exploration, debugging output, and excessive permission requests. Appendix D provides useful cases, but they do not establish the prevalence or relative contribution of these causes. Consequently, it remains difficult to determine which safety weaknesses dominate the aggregate results and what safety scope guidance specifically improves. This limits the paper’s ability to provide deeper insights and actionable guidance beyond documenting the observed phenomenon.

  2. The incremental value of the construction pipeline and the reliability of the final evaluation are insufficiently validated. Workflow Analyzer combines static workflow recovery, multiple construction roles, and iterative review, but it is not compared with a simpler pipeline in which an LLM directly analyzes source code and constructs tests. The quality or efficiency gains attributable to these additional components therefore remain unclear. Furthermore, the human audit in Appendix C identified grading errors, verifier coverage gaps, and ambiguous task contracts. Although the authors describe subsequent revisions, they do not report sufficiently comprehensive results from a final re-audit, and the two annotators reviewed disjoint task sets. Because the main empirical conclusions depend on these task boundaries and verifiers, clearer evidence is needed that the evaluation accepts legitimate alternative solutions and reliably detects violations across different execution strategies.

Questions for authors
  1. How much do static workflow recovery and iterative review improve the quality of the constructed tasks? A comparison with a simpler approach that directly asks an LLM to read the skill code and generate tests would help clarify their contribution.

  2. How were the five fixture categories chosen? Please report the number of tasks and the safety/UFR results for each category, accounting for tasks that involve multiple categories.

  3. What are the main causes of the violations detected during execution? Beyond the breakdown in Figure 3, can the authors distinguish failures arising from skill guidance from those introduced by the agent itself, and explain which failures SG reduces?

  4. How were the problems identified in the human audit corrected and rechecked? Was a subset of tasks independently reviewed, or tested with alternative solutions, to check whether the verifiers reject valid behavior or miss violations?

  5. Why was GPT-6 Astra used to construct the benchmark but not included in the evaluation? It would also help to know whether the reported SG improvements hold across repeated runs, with uncertainty estimates for the paired comparisons.

Comments on the paper

1 comment
Area_Chair_PdJSOpenReview ↗

Please engage in the rebuttal stage

Dear Reviewers,

Thank you for your time and thoughtful reviews. As the discussion period draws to a close, I would appreciate your consideration of the authors’ responses and any new evidence. Please share any remaining concerns and update your reviews and scores where appropriate. Thank you for your constructive participation.

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.