acceptodds
Under review as a conference paper at ICLR 2027

SkillGovernor: Localizing and Governing Risk across Agent Skill Graphs

Abstract

LLM agents accomplish complex real-world tasks by loading skills, which turns the skill package into an exploitable control surface. Attackers manipulate the content, structure, and interrelationships of skill files to inject payloads, so the skill still appears benign while the payloads couple with the procedures the agent depends on. Existing audit methods locate risk without modeling how it propagates across the skill's structure, and runtime guardrails block risky actions while leaving the risk-bearing content in place. We propose SkillGovernor, an end-to-end framework that governs a risk-bearing skill at its source. Its Risk Audit organizes the skill into a graph, reasons over it in cascade, and joins scattered risk into a risk-connected view. Its Dynamic Governor operates on that view and chooses the least intervention among graded operations, one that removes the risk and keeps the skill's legitimate function. We evaluate on 78 cases from two public skill-security benchmarks, across five configurations on three agent systems. SkillGovernor reduces the attack success rate (ASR) by 26 percentage points on average while maintaining task completion, reaching the lowest or tied-lowest ASR in every configuration. Its structure-based audit raises the injection point detection rate from 0.26 for a naive judge to 0.76. Together, these results show that a skill's risk can be governed at its source without compromising functionality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agent, Agent Skill, Indirect Prompt Injection, Agent Security, Risk Governance

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RFiSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

SkillGovernor governs risky skills at their source before execution. A Risk Audit builds the skill’s dependency graph and reasons over it in a bounded cascade to produce a risk-connected view; a Dynamic Governor then applies graded operations (edit/rewrite/delete/alert) under dual verification. On 78 cases across five configurations, ASR drops 0.32→0.06 with task completion held, and injection-point detection rises 0.26→0.76.

Critical strengths
  1. Intervene-at-source framing. Rather than blocking actions at runtime or merely flagging, SkillGovernor removes risk-bearing content while preserving legitimate function. The least-intervention grading (edit vs rewrite vs delete) is a thoughtful middle ground that the delete-only and alert-only ablations justify.
  2. Strong, well-ablated audit. Graph structure plus bounded cascade lifts IDR to 0.76, clearly above SkillAudit (0.50) and a naive judge (0.26). The w/o-Cascade and w/o-Bound ablations show each mechanism contributes, and the contextual/obvious split is a genuinely informative cut.
  3. Honest, careful evaluation. A clean-skill control measures false positives (1.8%), Table 7 transparently lists both recovered and broken tasks, and the authors repeatedly flag that one case equals 1.3 points. Manual injection-point annotation and released prompts add credibility and reproducibility.
Critical weaknesses
  1. Margins sit at benchmark resolution. With 78 cases, one case moves rates 1.3 points. SkillGovernor’s 0.06 ASR versus ARGUS’s 0.09 is a 3-point gap — barely above noise — and the w/o-Scheduler ablation (0.06 vs 0.07) suggests one of the two core roles is nearly inconsequential.
  2. Governance both heals and harms. “Maintained TCR” masks Table 7: several done→failed cases are offset by failed→done ones. Net-zero is presented favorably, but it means source editing sometimes damages legitimate function, and the balance may not hold on other skill distributions.
  3. Coverage ceiling and residual risk. IDR of 0.76 means 24% of injection points are never surfaced, so a quarter of risk is never governed and still needs the runtime layer the paper doesn’t provide. Pre-execution governance is therefore a complement, not the standalone defense the framing suggests.
  4. Cost and model dependence. Audit, Scheduler, and Editor run on strong proprietary models (GPT-5.4, DeepSeek-V4), with median 150s and p90 917s per skill. Heavy reliance on frontier auxiliary models plus curated case selection (“most realistic”) raises cost and selection-bias concerns.
Questions for authors

see weakness. Moreover, I think there is a problem with the rendering of fig1.

2 comments

Thank you for recognizing the value of source-level intervention, the audit ablations, and the transparent reporting of task outcomes. We address the interpretation of the performance margins, individual task changes, residual risk, and deployment cost below.

W1 | Margins at benchmark resolution

We agree that the small ASR margins alone should not carry the argument. The two comparisons concern different contributions:

ARGUS. Our contribution is supported by the joint security-and-utility result. SkillGovernor and ARGUS achieved pooled ASRs of 5.6% and 8.7%, respectively. The 95% case-level interval for ARGUS minus SkillGovernor is [−0.8, +7.2] points, so we do not treat this as a definitive ASR ranking. Task completion separates the two more clearly: the average TCRs are 0.41 and 0.33, and the paired difference is 8.5 points higher for SkillGovernor, with a 95% case-level interval of [+4.6, +12.6]. In every configuration, SkillGovernor achieved the same or lower observed ASR and the same or higher observed TCR.

Scheduler. The ASR difference is small: 7% without the Scheduler and 6% with it. Its distinct function is to independently check the Editor's proposals and record the decisions. It accepted 96% of proposals and returned 8 nodes for revision, making these interventions traceable. The contribution of the full governance mechanism is demonstrated separately by the fixed-audit ablation: 6% ASR with the full Governor, compared with 11% for delete-only and 14% for alert-only. Thus, the evidence for governance does not rest on the Scheduler's one-point margin.

W2 | TCR changes in both directions

Our claim concerns aggregate task completion, and Table 8 explicitly reports both gains and losses. We agree that the average does not guarantee preservation of every task. Among the case-configuration pairs where governance blocked an otherwise successful attack, Table 8 records 9 done-to-failed transitions across 7 cases and 16 failed-to-done transitions across 8 cases.

The clean-skill control helps interpret these changes. Even among untouched cases, three tasks moved in each direction. Task outcomes can change between executions without any source edits, so a changed outcome alone cannot establish damage from editing.

The evidence supports maintaining aggregate task completion on the evaluated cases while substantially reducing ASR. We report the individual losses alongside the gains, and do not assume that their balance holds for every skill distribution.

W3 | Coverage ceiling

We agree that pre-execution governance is one layer, with runtime safeguards at a different intervention point. The audit missed 63 of 267 annotated injection points, which the Governor could not directly address. This 23.6% is a proportion of injection points, not an attack success rate. End-to-end evaluation separately measured an average ASR of approximately 6%, down from 32% without defense.

Runtime safeguards can additionally inspect attempted actions during execution. We evaluated pre-execution governance and runtime defenses separately, so the effectiveness of a combined pipeline has not been measured.

W4 | Cost, model dependence, and case selection

We agree that auxiliary-model cost and case selection affect deployment and generalization. Our evaluation establishes the feasibility of governing useful skills with the specified models and cases. The relevant costs and selection criteria are explicit:

Cost. The median governance time is 150.6 seconds per case and p90 is 917.4 seconds, over the 50 cases the undefended run attacked successfully. The full audit, graph construction plus judging, has a median of 226.9 seconds per case over the 78-case set. The work is performed offline, so its cost is amortized across repeated uses under unchanged skill content and task requirements.

Model dependence. We used GPT-5.4 for the audit and Scheduler based on preliminary safety checks, and DeepSeek-V4 for the Editor based on its task completion in the base runs. All compared methods used the same judge model. Whether smaller auxiliary models retain the same performance is not established.

Case selection. Appendix A requires each case to retain a meaningful task after injection removal and provide the assets needed to complete it. This criterion is necessary to evaluate functionality preservation. We also favored realistic workflows and reduced repetition among similar skill variants, yielding 52 SkillSafetyBench and 26 OpenSkillRisk cases. All methods were evaluated on the same curated set. Comparison holds within this setting, and generalization depends on the skill distribution.

Additional question | Figure 1 rendering

We have checked Figure 1 on multiple devices and with different viewers and have not reproduced the rendering issue. PDF and PNG versions are included in the supplementary materials.

Thank you for the response, which addresses my concern. I have updated the rating to 3.

Reviewer_AWu3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents SkillGovernor, a framework for mitigating prompt-injection risks in agent skills through a two-stage process consisting of risk auditing and automated remediation. The proposed approach models a skill as a dependency graph, enabling the identification of both directly malicious components and intermediary elements that facilitate the propagation of harmful instructions across the skill structure. Building on this analysis, SkillGovernor applies a hierarchy of remediation actions, including editing, rewriting, and deletion, with the objective of reducing security risks while preserving the skill's intended functionality. Empirical evaluation on benchmark datasets suggests that the approach can reduce attack success rates while maintaining comparable task performance. Overall, the paper's contribution lies in integrating graph-based risk localization with automated skill modification, providing a practical mechanism for sanitizing agent skills prior to deployment rather than relying solely on detection or runtime safeguards.

Critical strengths

Strength 1: End-to-end integration of risk localization and remediation

The paper addresses a practically important gap in the security of agent skills by combining risk identification and risk mitigation within a single framework. Rather than limiting itself to detecting prompt-injection risks or blocking unsafe behavior at runtime, SkillGovernor attempts to localize risk within the skill structure and then actively remove it through targeted modifications. This integrated formulation is consequential because it directly targets the skill artifact itself, offering a more complete defense pipeline than approaches that only provide warnings or runtime safeguards. The experimental results suggest that this design can substantially reduce attack success rates while largely preserving task completion, supporting the practical relevance of the proposed approach.

Strength 2: Structured graph-based reasoning for skill analysis

A key technical contribution is the representation of skills as dependency graphs and the use of graph-based auditing to identify both risk-bearing components and the pathways through which risk propagates. This formulation appears well aligned with the structure of modern agent skills, where malicious behavior may be distributed across multiple files and dependencies rather than concentrated in a single instruction. The ability to reason about interconnected components enables more precise localization of risks and provides a principled basis for subsequent remediation. This aspect is consequential to my assessment because it represents the paper's primary methodological innovation and is supported by empirical improvements in injection localization performance.

Critical weaknesses

Weakness 1: The effectiveness of the governance stage is fundamentally limited by incomplete risk localization

While the paper's primary contribution is the integration of localization and remediation, the reported benefits of the governance framework depend critically on the quality of the audit stage. The evaluation shows that the proposed audit identifies only approximately 76% of annotated injection points, implying that a non-trivial fraction of risks remain undiscovered and therefore cannot be governed. As a result, it is difficult to disentangle whether the reported gains stem from the governance mechanism itself or from the subset of attacks that happen to be successfully localized. This limitation is consequential because the paper's central claim is an end-to-end risk-remediation pipeline, yet the effectiveness of that pipeline appears bounded by the recall of the auditing component.

Weakness 2: Limited evidence that graph-based auditing is necessary or uniquely responsible for the observed improvements

The paper argues that representing skills as dependency graphs enables more precise localization of distributed risks and serves as the key methodological innovation. However, the empirical analysis provides limited evidence that the graph-based formulation itself is the primary driver of performance gains relative to simpler alternatives. While ablations demonstrate some benefit from the proposed auditing procedure, the paper does not fully isolate the contribution of graph-based reasoning from other factors, such as additional LLM-based analysis, verification stages, or increased inference-time computation. Consequently, it remains unclear whether the added complexity of constructing and reasoning over skill graphs is essential to achieving the reported results. This issue is consequential because the graph-based audit constitutes the paper's principal technical novelty and therefore warrants stronger evidence of its necessity and advantage.

These weaknesses are stronger than generic concerns about LLM judges because they directly target the two central claims: (i) that localization and remediation together provide an effective end-to-end solution, and (ii) that graph-based auditing is the key innovation enabling that solution.

Questions for authors

Q1: The proposed audit achieves an injection detection rate of 0.76, implying that approximately one quarter of annotated injection points remain undiscovered. Can the authors characterize the missed injection points (e.g., by attack type, skill structure, or dependency pattern) and provide evidence that they do not represent a systematic blind spot of the proposed approach?

Q2: The paper argues that graph-based auditing is a key contributor to improved localization performance. Can the authors provide further evidence that the observed gains arise from the graph-based representation itself, rather than from increased inference-time computation, additional model calls, or more extensive reasoning compared to the baselines?

Q3: Can the authors provide a breakdown of end-to-end failures, distinguishing between (i) risks missed during auditing, (ii) risks correctly localized but unsuccessfully governed, and (iii) failures of the verification stage?

Q4: The evaluation focuses on individual skill packages. To what extent do the authors expect the proposed auditing and governance procedures to remain effective when risks emerge through interactions among multiple skills or through skill composition?

Q5: How sensitive are the audit and governance stages to errors in the underlying LLM judgments used for risk assessment, functionality preservation, and verification? Are there analyses measuring the stability of results across different models or auditing configurations?

4 comments

Thank you for evaluating our two central claims separately. Your questions address the respective contributions of risk localization and governance, which is precisely the distinction our experiments are designed to investigate.

W1 | Separating the contribution of governance from the effect of audit coverage

We agree that incomplete audit coverage leaves risks unaddressed. However, this does not prevent us from identifying the Governor's contribution. Our end-to-end ASR is computed over all 78 evaluation cases, including cases with missed injections. The reported gains therefore already reflect the consequences of incomplete localization.

More directly, the existing ablations in Section 4.3 evaluate the same cases with the same audit output and vary only the intervention. ASR is 14% with alert-only, 11% with delete-only, and 6% with the full Governor. Because the localized risks are held fixed, differences in which attacks the audit detects cannot explain these gains. This controlled comparison demonstrates that the governance mechanism contributes beyond localization.

The intervention determines what happens to the located risk. Alert-only retains the risky content. Delete-only removes the identified span, but a risk-bearing expression can extend into the surrounding content. The full Governor applies graded operations that account for this coupling, reducing the residual attack success left by these simpler interventions.

Audit coverage constrains which risks can be addressed, while the Governor determines how effectively the identified risks are mitigated. Our experiments establish this contribution under the audit's actual, imperfect coverage. Q3 provides the breakdown of remaining end-to-end failures.

Q1 | Missed injection points

We have characterized the 63 missed injection points (23.6% of the 267 annotated points). The largest single category accounts for 33% of the missed points and 7.9% of all annotated points, and misses appear across attack types, file types, structures, and span lengths. All percentages in the table use the full set of 267 annotated points as the denominator.

DimensionMissed-point distribution
Attack typePoisoning 7.9%, Bias Manipulation 7.1%, External State Manipulation 3.4%, each remaining type at most 1.1%
File typeMarkdown 14.6%, scripts 4.1%, configuration 3.7%, other files 1.1%
StructureSingle node 21.0%, multiple nodes 2.6%
Span length1 line 6.4%, 2–3 lines 3.7%, 4–8 lines 7.5%, more than 8 lines 6.0%

No misses were recorded for Persistence Implantation, Data Harvesting, or Control-Plane Hijacking. Poisoning and Bias Manipulation account for 63.5% of the 63 missed points. Using each attack category's own total as the denominator, External State Manipulation has the highest miss rate at 41%.

W2 / Q2 | Graph vs. computation

The existing ablation shows that additional model calls cannot explain the cascade's improvement. With the judge and call count fixed, independent per-node judgments achieved an IDR of 0.62, while propagating judgments along dependencies achieved 0.76 (Section 4.4). This comparison separates the benefit of dependency-aware propagation from an increase in call count.

Two further comparisons clarify the mechanism. Supplying the entire graph in each cascading judgment achieved 0.69, supporting bounded context selection. Separately, the whole-skill baseline without a graph achieved 0.26. Together, these results are consistent with structured decomposition, dependency-aware propagation, and bounded context selection. The cascade ablation fixes call count while the input contexts differ. The w/o Bound variant supplies the whole graph in each judgment, so its inputs are longer, and it still reaches only 0.69. More context per judgment does not recover the gain.

Q3 | Breakdown of end-to-end failures

In the evaluated runs, all 11 skills involved in end-to-end failures had audit localization errors. We observed no failures attributable to unsuccessful governance after correct localization or to verification. Following your three categories, the breakdown is:

(i) Failures attributable to missed detections by the audit: Of the 11 skills, the audit missed all injection points in 3 skills, missed part of the injection points in 7 skills, and imprecisely marked the injected content in 1 skill.

(ii) and (iii) No observed cases. Neither the governance stage nor the verification stage was identified as the cause of an end-to-end failure in this evaluation.

Q4 | Multi-skill composition

We expect the approach to extend to multi-skill workflows whose cross-skill dependencies can be represented explicitly. Graph-based dependency modeling gives these interactions a natural representation. It makes the links between instructions and scripts explicit, the starting point for tracing risk across skill boundaries. For example, if skill A invokes an injected script in skill B, a joint graph could connect the risky script to the invocation that activates it. The audit could then propagate the risk judgment along that dependency and guide intervention on the implicated content. This extends the reasoning already applied across files within a skill. It would require constructing graphs across skills and verifying that interventions preserve the combined workflow's legitimate function.

Coverage would be more limited when harmful interactions cannot be inferred from the available artifacts, such as dynamically selected skills or previously unobserved shared-state dependencies. Those interactions need execution context or traces before the audit can see them. The key condition is whether the joint representation captures the interactions that cause the risk. A risk that arises from a capability union rather than from any single skill is harder still, since it has no single node to attribute it to.

Our current implementation and evaluation cover individual skill directories, as stated in Appendix E, and both cases remain to be tested in composed workflows.

Q5 | Sensitivity to LLM judgment errors and cross-model stability

The question concerns three stages and one stability question. The stages fail in different places: the audit decides which risks reach governance, the editor determines repair quality, and the scheduler decides what is preserved and whether a change is accepted. We answer the three stages from measurements we already have, and state what remains open on stability.

Risk assessment. The audit's error is measured directly: on clean skills, 59 of 3,270 audited nodes were false positives (1.8%). Configuration changes this stage measurably: without the cascade, IDR falls from 0.76 to 0.62, and contextual IDR from 0.74 to 0.47. Without the bound, it falls to 0.69. The errors that remain are visible downstream: every end-to-end failure we observed traces to a localization gap (Q3).

Functionality preservation. The operation set constrains damage before any model judgment: non-terminal operations must preserve the node's legitimate function F, and a node is deleted only when F is empty. In the 48-case clean-skill control, completed tasks changed from 19 to 20 after governance, so aggregate utility was not visibly reduced on uninjected skills in this evaluation.

Verification. The Scheduler checks every proposal and records each verdict. It accepted 96% and returned 8 nodes for revision, so a wrong semantic judgment can be reviewed and corrected, and mechanical checks enforce edit boundaries, parsing, and reference integrity. In the observed runs, verification caused no end-to-end failure (Q3).

Stability. Across the five downstream configurations, spanning four models and three agent systems, ASR falls and task completion is maintained. The audit and Scheduler are fixed to GPT-5.4 and the Editor to DeepSeek-V4 throughout, and the baselines use the same judge, so the comparison is not confounded by model strength. Replacing the auxiliary models would primarily evaluate the safety sensitivity of the models themselves. Our focus is on what the framework contributes with the model held fixed: a more favorable overall balance between attack suppression and task completion among the methods evaluated here. Model-swap sensitivity remains unmeasured in our experiments.

Thanks for the detailed response. Most of my questions and concerns have been addressed. I have one last concern, though. It seems like your results suggest that real bottleneck for this problem is the audit not the governor and, hence, future works should focus more on solving the audit problem.

We sincerely thank you for the update and for the careful reading behind it. We agree that localization is the limiting factor in our results. The audit determines which risks reach governance, and every residual end-to-end failure we traced came from a localization gap. Improving audit coverage is therefore the most direct way to raise end-to-end performance, and we see it as the most promising direction for future work.

The Governor is where a localization result becomes a targeted operation. It decides what happens at each place the audit surfaces, removing the risk while preserving legitimate function. Among all baselines, SkillGovernor achieves the best overall balance, with the largest ASR reduction, 26.9 points, and task completion at the undefended level. With the audit fixed, the choice of operation still matters. ASR is 14% with alert-only and 11% with delete-only, against 6% with the full Governor. Localization sets the ceiling, and the Governor determines how much of it the pipeline reaches.

Reviewer_1dKfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces SkillGovernor, a framework that aims to guard against risky skills by first reasoning over the skill by representing it as a graph (of sections or code blocks). Then they use an auditing judge to reason over the graph, to capture how the risk travels over nodes in the skill. Then a Dynamic Governor reasons over the identified risk to select an intervention to remove the risky parts while keeping the skill's legitimate function. They evaluate on cases selected from skill benchmarks whose skills remain useful after injection removal, showing reduced attack success rate while average task completion rate remains high.

Critical strengths
  • Method studies the skill risk detection problem in a way that differs in some ways from existing work, e.g. using the graph-based reasoning approach, and attempting to intervene to remove the risky parts while keeping the legitimate function.
  • Experiments show good results on five backbone models, including on both ASR and task completion, and under different conditions.
Critical weaknesses
  • Sample size tends to be on the small side (78 cases for table 1, 39 cases for table 2), making the results less convincing particularly considering that the results on ARGUS are in some cases close to the proposed method. Taking an average over 3 attempts might help to alleviate this somewhat.
  • The use of graphs seems like it should have an ablation to demonstrate that it is actually beneficial compared to a more straightforward approach of using LLMs directly on the entire skill (since they already have some inherent ability to reason over large interconnected codebases).

A weaker concern is that since the goal of removing the injected part of skills while keeping the legitimate parts is one of the more unique aspects of the method, it might be good to have a bit more motivation for this, e.g. one might argue that if a skill contains an injection, this already shows that its author is malicious or untrustworthy and it seems unlikely in practice that using skills from such authors would be practically useful? However, I am not familiar with the real world case, so I am open to rethinking on this point.

Questions for authors

See weaknesses

2 comments

We thank you for recognizing the methodological contribution and the results across five configurations. We address the requested ablation first, followed by repeated runs and the motivation for selective governance.

W2 | Ablating the graph against directly applying an LLM to the entire skill

We have already evaluated the requested whole-skill LLM baseline in Section 4.4. It reads the entire skill in one pass without constructing a graph and achieves an IDR of 0.26, compared with 0.76 for the full audit. The accompanying ablations examine propagation and context selection:

ComparisonSettingIDR
naive judgewhole skill in one pass, no graph0.26
w/o Cascadenodes judged independently, same number of calls0.62
w/o Boundcascade kept, whole graph in each judgment0.69
Full auditcascade under the bound0.76

With the judge and call count fixed, propagating judgments along dependency relationships increased IDR from 0.62 to 0.76. Keeping the cascade but supplying the entire graph in each judgment achieved 0.69, supporting bounded context selection. These comparisons identify benefits from specific mechanisms operating on the graph, alongside the comparison with the whole-skill baseline.

W1 | Small sample, repeated runs

We have not repeated each configuration three times because rerunning all 78 cases across five configurations would require substantial additional execution time and LLM tokens. Your concern is particularly relevant to the small ASR difference from ARGUS. Our contribution does not depend on treating that difference as a decisive ranking.

Table 1 provides the joint security-and-utility evidence: SkillGovernor achieved an average TCR of 0.41 versus 0.33 for ARGUS, with the same or lower observed ASR and the same or higher observed TCR in every configuration. This result supports the value of preserving legitimate task utility alongside strong attack suppression.

ComparisonΔASR (pp) ↓, 95% CIΔTCR (pp) ↑, 95% CI
Gov − Base−26.9 [−33.1, −20.8]+1.8 [−1.5, +5.9]
Gov − ARGUS−3.1 [−7.2, +0.8]+8.5 [+4.6, +12.6]

Differences are Gov minus the reference in each row, with 95% case-level intervals.

The interval for the 3.1-point ASR difference from ARGUS contains zero. The improvement over the undefended baseline is much larger: a 26.9-point reduction on average, with the confidence bound closest to zero indicating a 20.8-point reduction. The observed reduction holds across all five configurations, spanning four models and three agent systems.

The TCR column gives the same comparison in the utility direction. Relative to the undefended baseline, TCR is maintained (+1.8 points, [−1.5, +5.9]). Relative to ARGUS, TCR is 8.5 points higher ([+4.6, +12.6]), with ASR never higher in any configuration. The joint result therefore rests on a measurable utility difference rather than on the ASR margin.

Repeated executions would measure run-to-run variability, which these case-level intervals do not estimate. We consider the existing joint results and controlled component ablations sufficient to support the main empirical contribution: targeted governance substantially reduces observed attack success while retaining useful task functionality. Establishing a stable ASR ranking over ARGUS is not required for that contribution.

Additional concern | Why govern a risky skill instead of rejecting it?

We agree that rejecting risky skills outright is viable at deployment. However, the boundary between legitimate and risky behavior is often unclear and depends on the task and granted permissions. The author is not necessarily at fault. Skills are distributed through third-party marketplaces and sharing channels, where poisoning can happen before they reach users. Discarding every skill that carries an injection would also throw away work whose author did nothing wrong. We study how models can localize and mitigate risks while preserving legitimate functionality, with the goal of evaluating and improving their ability to perform precise skill governance.

For example, writing project files may be appropriate for an authorized editing task but exceed the permissions needed for read-only analysis. Our formulation conditions risk assessment on the task, user assets, and risk rules. The experiments study this capability using injected skills that still serve legitimate tasks after injection removal.

The fixed-audit ablation demonstrates why the intervention itself matters: the full Governor achieved 6% ASR, compared with 11% for deleting identified risk spans and 14% for issuing alerts. Rejecting an untrusted source remains a deployment option. Our research evaluates the ability to make targeted changes when retaining useful functionality is part of the objective.

I thank the authors for the clarifications and added experiments. I maintain my positive score.

Comments on the paper

2 comments

Overall Response (to the Three Reviewers and the AC)

We thank the reviewers for recognizing the value of combining risk localization with targeted remediation of skill artifacts. The main concerns are whether governance contributes beyond detection, whether dependency-aware auditing improves localization, and whether the evaluation supports preserving task utility alongside security. Three results from the evaluation address these concerns:

  • Governance contributes beyond detection. With the audit output fixed, ASR (attack success rate) is 14% with alert-only, 11% with delete-only, and 6% with the full Governor (Section 4.3).
  • Dependency-aware propagation improves localization. With the judge and call count fixed, propagating judgments along dependencies increases IDR (injection point detection rate) from 0.62 to 0.76. Separately, the requested whole-skill, single-pass baseline achieves 0.26 (Section 4.4).
  • The contribution extends beyond the small ASR margin over ARGUS. SkillGovernor achieves an average TCR (task completion rate) of 0.41 versus 0.33 for ARGUS, with the same or lower observed ASR and the same or higher observed TCR in every configuration (Table 1).

Below, we explain these comparisons, report the requested error analyses, and clarify the scope of the evaluation.

Following the reviews, we have revised the paper to make the evidence behind these responses easier to check:

  • Section 4.2 now states the benchmark's resolution, reports the case-level intervals for the ASR comparison with ARGUS and the paired task-completion difference, and points to the individual task-completion transitions in Table 8.
  • Section 4.4 now separates the detection gap from attack success and reports the stage attribution of the residual end-to-end failures.
  • Section 4.1 now states the case retention criterion and notes that the judge model is shared across all compared methods.
  • Appendix C now reports the distribution of the missed injection points.

The revised PDF marks these changes in blue.

Area_Chair_2Q8KOpenReview ↗

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.