acceptodds
Under review as a conference paper at ICLR 2027

Hiding in Plain Sight: Decoupling Pretext from Actuation for Skill Poisoning in LLM Agents

Abstract

LLM agents increasingly rely on reusable Skills for complex, multi-step tasks, creating a critical supply-chain attack surface where poisoned Skill content steers agent decision loops under benign requests. Existing skill poisoning attacks either colocate actuation with its contextual pretext or distribute actuation across multiple Skills, but do not explicitly separate the rationale for execution from the operation itself. In this work, we reveal that untrusted agent decisions fundamentally depend on two conceptually distinct Risk-Realization Factors (RRFs): an actuation factor, specifying what concrete operation is performed, and a pretext factor, providing the situational rationale for why the agent must perform it. Guided by this abstraction, we propose a coordination-based attack paradigm: decoupling pretext from actuation. Rather than fragmenting the malicious actuation, we preserve it as an intact operation within a downstream Steering Skill, while delegating the pretext factor to an upstream Grounding Skill that subtly alters persistent environment artifacts through routine utility operations. The intact actuation thus hides in plain sight, appearing completely legitimate and task-driven only when evaluated against the fabricated pretext. Building on this formulation, we develop an automated framework that discovers authentic execution dependencies, synthesizes coordinated pretext–actuation skill pairs, and iteratively refines poisoned skill instructions via runtime closed-loop feedback. Extensive evaluations across single-session and persistent cross-lifecycle scenarios demonstrate that decoupled skill poisoning achieves high attack success, exposing a critical blind spot in isolated Skill security audits. Our automated framework code is available at https://anonymous.4open.science/r/CoordPoison-C2AB/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent Security, Agent Skill, Skill Poisoning

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_UQv7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates supply-chain "skill poisoning" of LLM agents, in which malicious SKILL.md instruction files steer an agent's decision loop during benign user requests. The main conceptual idea is to decompose a payload's realization into two Risk-Realization Factors (RRFs): an actuation factor A(P)\mathcal{A}(P) (the concrete harmful operation, or what is executed) and a pretext factor Z\mathcal{Z} (the situational rationale, or why it looks warranted) (Sec. 3.1, Eq. 1). Rather than fragmenting the actuation across skills as in prior distributed attacks, the proposed approach keeps A(P)\mathcal{A}(P) intact inside a downstream "Steering Skill" and offloads the pretext to an upstream "Grounding Skill" that writes a persistent workspace artifact aZa_Z, which is later read as justification (Sec. 3.2, Eqs. 2–3). The attack operates strictly over authentic, pre-existing benign coordination between two skills (Sec. 3.3). The authors implement this in CoordPoison, an automated pipeline featuring dependency screening, plan-then-render decoupling, and failure-guided refinement up to Nmax⁡=12N_{\max}=12, using GPT-5.5 for generation (Sec. 4). They evaluate the method on 62 qualified skill pairs gathered from public GitHub skill repositories, paired with 7 fixed payloads adapted from the SkillJect and Skill-Inject benchmarks (434 instances total), across three primary victim models. The experiments also cover cross-model transfer (MiniMax-M3, GPT-5.5), ablations, a cross-lifecycle analysis, and a prompt-level defense (Sec. 5, Tables 1–4, 8).

Critical strengths

The attack transfers, and the cross-session effect is explained. Attacks built on one model reach roughly 79–95% ASR on a held-out model (MiniMax-M3) vs. 0% for the naive baseline (Table 2), and splitting the attack across sessions raises success further (e.g., DeepSeek 68.25% → 95.31%; Table 3).

Critical weaknesses
  • W1. Headline within-model efficacy numbers are hard to interpret and potentially optimistic, with inconsistent metric reporting and no variance. Table 1 is titled "Generation-stage attack performance," and Figure 3 shows that its values are the cumulative fraction of instances resolved by the optimizer over up to 12 refinement iterations (the endpoints 76.2% / 51.2% / 44.4% match Table 1's overall ASRs). This makes it unclear whether Table 1 reports in-sample optimization success against each victim or a held-out re-evaluation of the finalized poisoned skills; both the title and Figure 3 point to the former. Three other issues compound this: (a) the column is labeled "ASR" (caption: successes "over eligible Steering-only failures"), but the text calls the exact same values "conditional ASR"/"cASR" (Sec. 5.2, lines 317–321), and ASR and cASR can differ substantially elsewhere (e.g., GLM cross-lifecycle is 60.58% ASR vs 91.30% cASR in Table 3), so knowing which metric is reported matters; (b) identical configurations produce different numbers across tables, as DeepSeek overall is 76.19 in Table 1 while Same-Lifecycle CoordPoison is 68.25 ASR / 69.35 cASR in Table 3, with no explanation of the denominators; and (c) no repeated runs, seeds, or confidence intervals are reported for ASR/cASR (the only N=3N=3 is for benign-baseline task-completion gating, Appendix A.1, ~line 621), despite stochastic LLM sampling. Because these are the core efficacy claims of the paper, their interpretation depends on whether they reflect optimizer success or deployment success, which metric is actually shown, and how noisy the measurements are. To resolve this, the authors should state clearly what each table measures, reconcile the denominators across Tables 1–3, and add a held-out re-evaluation of the finalized poisoned skills across multiple runs with variance.
  • W2. The specific novelty of decoupling pretext over other multi-skill attacks and the "evades isolated audits" claim are argued but not empirically demonstrated. All three baselines (Naive, Skill-Inject, SkillJect) are localized single-skill attacks (Sec. 5.3; Appendix C.3), and the "w/o Pretext" ablation (Table 3) is also a single-skill variant. The held-out transfer table (Table 2) does show CoordPoison beating these single-skill baselines (e.g., 95.19% vs 74.04% for SkillJect on MiniMax for GLM-source), which supports the view that multi-skill coordination helps. However, it does not test the paper's more specific claim that decoupling pretext beats the conceptually nearest attacks that fragment actuation (SkillTrojan, ColluSkill) or exploit persistent cross-lifecycle state (SkillHarm, SkillJack); those are only discussed in related work (Sec. 2.2) and never compared empirically. Along similar lines, the central claim of exposing a "blind spot in isolated Skill audits" (Abstract, line 32; Conclusion, lines 484–485) is only tested against a prompt-level runtime auditing defense (Table 4), without running any concrete single-skill scanner or auditor on the poisoned pairs to show they actually pass. For an attack paper, isolating the advantage of the proposed mechanism and showing the claimed evasion in practice are central to the contribution. The authors should run a concrete single-skill audit tool on the poisoned pairs (which requires only their existing artifacts and should be feasible) and, where possible, compare against at least one representative multi-skill or persistent-state attack. (p.s., I note that these competing attacks are all 2026 preprints and likely contemporaneous with this submission, so a head-to-head comparison is a strengthening suggestion rather than a requirement; the scanner evaluation is the more actionable request.)
Questions for authors

Please refer to the weaknesses.

Reviewer_mcRMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2(+1)
Read review
Paper summary

This paper presents CoordPoison, an automated skill-poisoning framework for LLM agents that separates the contextual justification for executing a malicious operation (pretext) from the operation itself (actuation). Unlike approaches that distribute malicious payload fragments across multiple skills, CoordPoison preserves the complete payload in a downstream Steering Skill while using an upstream Grounding Skill to manipulate intermediate workspace artifacts. The resulting artifact is intended to provide a plausible justification for the subsequent malicious execution.

The framework consists of three main stages: (1) identifying authentic execution dependencies between benign skills and screening out instances vulnerable to Steering-only poisoning; (2) constructing coordinated Grounding–Steering skill pairs; and (3) iteratively refining attack instructions using runtime execution feedback.

The authors evaluate the approach using 62 qualified skill pairs and seven payloads across several LLMs. The experiments investigate attack success, cross-model transferability, component ablations, cross-lifecycle persistence, and prompt-level defenses.

Critical strengths

S1. Technical novelty and innovation.

The distinction between contextual pretext and executable actuation provides a useful perspective on compositional skill-poisoning attacks. In particular, preserving the malicious operation in one skill while manipulating its execution context through another is conceptually different from explicitly splitting malicious payloads across skills.

S2. Significance of the investigated problem.

The work addresses a relevant security concern in skill-enabled LLM agents: untrusted instructions can influence execution not only directly but also indirectly through persistent workspace state. Understanding such interactions may inform future research on provenance-aware and composition-aware security mechanisms.

Critical weaknesses

W1. The central causal claim is not sufficiently established.

The paper claims that the upstream Grounding Skill creates contextual legitimacy that causes the downstream agent to execute an otherwise rejected malicious operation. However, the experiments do not adequately distinguish a genuine change in the downstream execution decision from an increase in the probability that the Steering Skill or payload is encountered.

This distinction is particularly important given Table 8. On DeepSeek-V4-Flash, the L2-only setting achieves only 14.06% ASR but already reaches 94.74% cASR. The cross-lifecycle setting achieves 95.31% ASR and 100.00% cASR.

These results suggest that much of the improvement may arise from increased actuation exposure or workflow reachability rather than a fundamentally different decision mechanism.

Moreover, observing that the Grounding Skill writes an artifact, the Steering Skill reads it, and the payload is subsequently executed establishes an execution sequence, not necessarily a causal dependency.

This substantially weakens the paper's primary mechanistic contribution.

W2. The baseline comparisons are not sufficiently controlled.

CoordPoison benefits from iterative runtime feedback with up to 12 refinement iterations. In contrast, Appendix C.3 explicitly states that the localized baselines do not undergo source-model optimization iterations.

Consequently, the observed improvements may reflect differences in optimization budget and access to runtime feedback rather than the benefits of pretext–actuation decoupling.

Additionally, the main attack evaluation conditions on Steering-only baseline failures. Such filtering produces different eligible subsets for different victim models, which complicates comparisons and limits the interpretation of the reported attack success rates.

Importantly, conditioning on Steering-only failure under the source model does not imply the corresponding baseline will fail under the target model. Indeed, Table 2 shows that the SkillJect baseline achieves 93.75% transfer ASR on GPT-5.5 for the DeepSeek-V4-Flash-selected subset, compared with 96.67% for CoordPoison.

The relatively small difference in this setting raises questions about the practical magnitude of the improvement.

Questions for authors

Q1. Causal mechanism. Given that L2-only achieves 94.74% cASR on DeepSeek-V4-Flash, how do the authors distinguish the effect of the upstream pretext on the downstream execution decision from its effect on skill invocation and payload exposure? Can the authors provide a matched counterfactual experiment that changes only the pretext-bearing artifact content?

Q2. Baseline fairness. Why are the localized baselines evaluated without the same iterative optimization budget used by CoordPoison? How do the results change when SkillJect and other strong baselines receive an equivalent number of runtime feedback and refinement iterations?

Q3. Attack success verification. For each payload category, what fraction of reported successes corresponds to attempted script invocation, completed script execution, and independently verified harmful effects? Do the conclusions remain valid when success requires the intended side effect?

Q4. Experimental consistency. Can the authors correct the GLM-5.3-Flash values in Table 1 and explain why the "w/o Pretext (SkillJect)" ablation in Table 3 produces nonzero ASR despite the Steering-only failure screening? Please also clarify the different success rates reported in Tables 1 and 3.

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

The paper proposes a novel method for skill-poisoning attacks: instead of fragmenting the malicious actuation, keep the actuation intact and decouple the pretext into an upstream “Grounding” skill to write a workspace artifact. Results show high conditional ASR, strong cross-model transfer, higher success in cross-lifecycle settings, and limited effectiveness of prompt-level defenses.

Critical strengths

interesting problem and approach, Automated pipeline with runtime feedback, with good empirical attack results

Critical weaknesses

ASR is evaluated only on pair payload instances where the steering baseline fails. This is a reasonable way to isolate the coordination effect, but it makes the headline numbers non-comparable to standard ASR and can inflate the attack efficacy.

There are no confidence intervals for ASR, no clustering by skill pair, and some cells are small

the current evaluation overstates the attack’s efficacy through conditional ASR, and does not compare against composition-aware defenses.

Questions for authors

What is the unconditional ASR over all qualified pairs, including steering-only successes? What fraction of pairs are screened out?

Can you report clustered confidence intervals by skill pair and per-pair success distributions?

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We summarize the recognized strengths and our responses to the main concerns.

1. Key Strengths Recognized by the Reviewers

  • Technical novelty and significance. Reviewer mcRM recognized the conceptual novelty of separating execution justification from harmful action and the significance of security risks arising from persistent workspace state.
  • Automated methodology. Reviewer AQoz highlighted the interesting research problem and automated attack construction with runtime feedback.
  • Empirical effectiveness. Reviewer AQoz recognized the strong results, while Reviewer UQv7 highlighted cross-model transferability and cross-lifecycle effectiveness.

2. Substantive Responses to the Major Concerns

Attack effectiveness and reliability (AQoz, UQv7, mcRM). Our original evaluation already provides complementary evidence through construction success (Table 1), independent execution of finalized attacks (Table 3), and cross-model transfer (Table 2). As clarified in our rebuttal, Steering-only screening excludes instances where strong localized poisoning succeeds, preventing these successes from being incorrectly attributed to coordination and overstating its effectiveness. We further clarified ASR/cASR and evaluation protocols, and provided full-set pipeline analysis, repeated-run statistics, confidence intervals, and per-pair distributions. These results strengthen the interpretation and reliability of our original findings.

Mechanistic evidence and causal attribution (mcRM, AQoz). Our original experiments already support the contribution of execution justification. Table 3 shows substantial ASR reductions when justification components are removed while preserving the harmful action: 68.25% to 10.32% on DeepSeek-V4-Flash and 52.38% to 30.95% on Claude-Sonnet-4.6. Sec. 4.2 verifies Grounding–Steering dependencies in successful execution traces, while Table 8 provides complementary lifecycle controls. We further conducted the requested matched artifact-content counterfactual, holding the downstream Skill, task, artifact path, and environment fixed. Valid pretext-bearing artifacts increase ASR by 37.50, 16.54, and 7.14 percentage points over neutral artifacts across three models. Together, the original evidence and supplementary counterfactual substantiate the contribution of execution justification to harmful-action invocation, beyond the mere presence of an artifact.

Automated construction and baseline comparison (mcRM). Our agent-driven framework automatically discovers authentic workflow dependencies and uses runtime feedback to diagnose and repair failed coordinated attack constructions. Refinement preserves the fixed harmful action while establishing and repairing execution justification across Skills. The original Table 2 already demonstrates effective cross-model transfer compared with the baselines. To address optimization-budget concerns, we additionally evaluated SkillJect with its original five-iteration refinement setting. CoordPoison maintains higher overall transfer ASR across all source-model settings: +15.38 points on GLM-5.3-Flash, +8.33 on Claude-Sonnet-4.6, and +1.36 on DeepSeek-V4-Flash. These results reinforce the original comparison. Moreover, CoordPoison's iterative refinement addresses the distinct challenge of discovering authentic workflow dependencies and constructing coordinated attacks through runtime feedback, enabling the systematic discovery and realization of previously underexplored attack opportunities in complex multi-Skill workflows. This agent-driven construction process is itself a central methodological contribution of our work.

Defense scope (AQoz, UQv7). Our original defense evaluation (Table 4, Appendix E) focuses on runtime security assessment. Following the reviewers' suggestions, we additionally tested a static Skill scanner and the composition-aware skil compose framework. These supplementary evaluations clarify the scope of our findings beyond the original study's focus on runtime execution decisions.

3. Positive Feedback Following the Rebuttal

Reviewer AQoz expressed satisfaction with the additional results and remained positive. Reviewer UQv7 explicitly confirmed that the main interpretability concern was resolved and maintained a positive assessment. Reviewer mcRM acknowledged that our responses meaningfully addressed several concerns and raised the Soundness assessment.

The reviewers' recognition of CoordPoison's strengths, together with the original evidence and supplementary results, provides strong support for its novelty, significance, and methodological contributions. We respectfully hope the Area Chair will consider these points in the final decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.