acceptodds
Withdrawn from ICLR 2027

Runaway Reaction: When Benign Skills Compose into Malicious Behavior

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agent skills package task-specific knowledge and procedures that can be composed to support complex agent tasks, while public marketplaces provide a growing pool of reusable skills. Existing security vetting, however, largely evaluates skills in isolation, leaving composition-induced risks underexplored. Such risks arise because composing benign skills expands the agent’s capability space, enabling behaviors unavailable to any skill alone. Interestingly, we find that directly composing benign skills can already induce malicious behaviors, even when every individual skill passes security vetting. We further find that some target malicious behaviors remain difficult to realize through direct composition, even when the selected skills collectively provide the required capabilities. To systematically evaluate this vulnerability, we present Compositional Risk Induction via Multi-skill Execution (CRIME), a framework consisting of three components. CRIME first uses the Malicious Plot Casting module to decompose a target malicious behavior into complementary requirements and identify suitable skill com positions. For initially ineffective compositions, the Runaway Reaction Steering module uses execution feedback to refine the selected skills toward the target while preserving their standalone benignity. The resulting composition is then passed to the Skill Reaction Chamber module, where the skill pair is executed in a sandbox to determine whether the target behavior has occurred, with unsuccessful cases returned for further refinement. Furthermore, we construct a benchmark of 4,000 public skills across eight cybersecurity behaviors for systematic evaluation of composition-induced vulnerabilities. Experiments across multiple LLMs reveal malicious behaviors in both direct and optimized compositions, showing that individual skill inspection is insufficient to capture multi-skill security risks.

Primary area:
generative models
Keywords:
Agent Skills, Skill Composition, AI Agent Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fSj2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether skills that pass individual security vetting can produce malicious outcomes when used together. CRIME retrieves complementary skill pairs from a corpus of 4,000 public skills using Malicious Plot Casting, tests their joint execution in a sandbox using the Skill Reaction Chamber, and, when necessary, refines skill instructions through Runaway Reaction Steering. The evaluation covers 560 selected pairs across eight cybersecurity behaviors and distinguishes DCA from ACA. Under facilitated prompts, Table 1 reports average success rates increasing from 27.9% to 73.8% for DeepSeek-V4-Flash and from 9.7% to 32.1% for GPT-4o-mini when both skills are optimized.

Critical strengths
  1. The paper proposes a framework that turns benign public skills into malicious compositions through capability-driven skill discovery and execution-guided optimization.
Critical weaknesses
  1. The benign individually, harmful together is already established [1,2,3]. Execution-based evaluation is also already present in prior work. The value of this paper is only in automated discovery plus refinement constrained by per-skill vetting. Moreover, the paper’s set-containment theorems do little to establish the distinction to prior works.

  2. The detailed ablations use two weak backbones, and there is no newer OpenAI-family or Anthropic-family comparison. Older models can establish that a vulnerability exists, but cannot establish its prevalence across prevailing agents.

[1] When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems

[2] STAC: When Innocent Tools Form Dangerous Chains for LLM Agents

[3] Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems

Questions for authors
  1. A model might fail because it cannot execute the chain, or because it refuses. Figure 3 does not separate these explanations, it conflates capability and safety.

  2. This paper is poorly writtened. What do the references to Table 4 and Table 4c refer to? It is also hard to capture the main contributions.

Reviewer_naxAOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Multiple individual AI intelligent body skills that appear normal and pass the security checks can also cause harm when deliberately combined by attackers. One skill is responsible for reading local configurations, and another is responsible for sending data externally. When used together, they may expose the keys in the configuration. The author proposed a method called CRIME, which searches for such combinations among 4,000 public skills; when direct combination fails, the content of the skills is adjusted based on the execution results, while still trying to ensure that each skill can pass the individual checks. In experiments on 8 types of risks such as data leakage and dangerous command execution, the author found that some skills can be combined to launch attacks without modification, and the attacks become more likely to succeed after adjustments. The article proves that merely checking whether each skill is secure is not enough; one also needs to consider the consequences when they are used together.

Critical strengths
  1. The paper focuses on the combined risks that are easily overlooked in individual skill security checks, and demonstrates that attackers, by leveraging the original functions of two skills, may cause consequences such as data leakage. This indicates that even if the skills themselves do not have obvious malicious instructions, the data flow and functional connection between the skills still deserve review, and have direct reference value for the current security audit of the skill market.

  2. The author runs skill combinations in a sandbox, determines whether the attack is successful based on whether the target hazard appears in the environment, and uses the execution feedback to continue adjusting the skills. The experiment also compared different models, prompt intensities, skill selection methods, and security review tools, which helps to determine which conditions are more likely to trigger combined risks, making the conclusion have more specific experimental support.

Critical weaknesses
  1. Some experimental results show inconsistent phenomena that cannot be explained. The skills required to be retained must undergo safety review, but the pass rate of SkillSafe in Table 2 is only 93.8%; the results in Figure 2 and Table 3 also have significant differences under strict benign prompts, and the authors need to explain the experimental conditions and statistical standards.

  2. The key implementation details are not adequately described. How to extract the skill capabilities, how to sort the candidate combinations, and how RRS calculates the progress of execution to guide modifications, the descriptions are rather general, which affects the understanding and reproducibility of the method.

Questions for authors

See in Weaknesses.

Reviewer_chVMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies security risks arising from the composition of individually vetted agent skills. It distinguishes Direct Composition Attacks (DCA), which combine unmodified public skills, from Adaptive Composition Attacks (ACA), which refine skill instructions using execution feedback. The proposed CRIME framework combines target-driven skill selection through Malicious Plot Casting (MPC), iterative refinement through Runaway Reaction Steering (RRS), and sandbox-based evaluation through the Skill Reaction Chamber (SRC). The evaluation uses a corpus of 4,000 public skills and 560 selected skill pairs spanning eight target behaviors. Under Facilitated prompts, Table 1 reports that optimizing both skills increases the mean attack success rate from 27.9% to 73.8% for DeepSeek-V4-Flash and from 9.7% to 32.1% for GPT-4o-mini. The paper also examines prompt conditions, backbone models, vetting engines, and skill-selection strategies.

Critical strengths

S1. The paper addresses a worthwhile security question.

The distinction between inspecting an individual skill and evaluating the consequences of joint execution is important. Separating unmodified compositions from adversarially refined compositions is also useful, because these settings involve different attacker capabilities.

S2. Evaluating externally observable environmental consequences is a strength.

SRC considers tool actions and state changes rather than relying only on whether the model produces a harmful-looking textual response. This is an appropriate direction for evaluating agent behavior.

S3. The experiments explore several relevant dimensions.

The evaluation examines which skill is optimized, how much guidance the task prompt provides, and how different vetting engines assess the resulting skills. These comparisons could support a useful empirical contribution once the evaluation conditions and controls are clarified.

Critical weaknesses

W1. The central “benign in isolation” premise is not sufficiently established, and the reported vetting results require clarification. The threat model states that each constituent passes standalone vetting and does not exhibit the target malicious behavior when executed in isolation (Section 3). RRS explicitly discards candidate compositions containing a skill that fails vetting, and Equation (11) imposes this as a hard constraint. Section 6.1 identifies SkillSafe as the constraint engine. However, Table 2 reports a mean SkillSafe pass rate of only 93.8% after optimization, with some categories reaching only 90.0%. The original skills also have a mean pass rate below 100%.

Under the same deterministic evaluator and acceptance criterion, a hard vetting gate should yield a 100% pass rate for the skills admitted through that gate. Scanner stochasticity, configuration changes, or differences in the evaluated populations could explain the discrepancy, but the manuscript does not establish which explanation applies. Please report the fraction of evaluated pairs for which the attack succeeds AND both constituent skills pass the stated vetting procedure. Marginal skill-level pass rates do not establish that the successful attacks satisfy the threat model.

More fundamentally, scanner acceptance, failure to complete a particular target in isolation, and behavioral benignity are different properties. The empirical evidence primarily establishes scanner acceptance, whereas the theoretical analysis assumes that the entire standalone capability space lies within the safety space. The paper should distinguish these claims and provide direct evidence for the stronger “benign in isolation” premise.

W2. Important results cannot currently be reconciled across figures and tables.

Figure 2(a), labeled Strict Benign, uses a radial scale with an upper limit of 6. Section 6.3 likewise emphasizes low success under this condition. In contrast, Section 6.6 explicitly states that Table 3 uses Strict Benign prompts, yet the MPC averages in that table are 17.0% for DCA and 52.3% for ACA Both. Different models, skill populations, or optimization settings could explain the difference, but these distinctions are not adequately specified.

Similarly, Table S1 reports 100.0% successful optimization for Credential Exfiltration, whereas Table 1 reports ACA Both success rates of 88.6% and 40.0% for the two primary models. These numbers are not necessarily contradictory, but the manuscript does not identify the configurations and denominators needed to relate them.

Please provide the exact model, prompt condition, evaluated population, denominator, and execution protocol for each result. In particular, distinguish success over all selected pairs, optimization success conditional on initial failure, success at least once within a search budget, and success on an independent execution of the final edited pair. These distinctions affect the central conclusions, especially the risk under benign task inputs. W3. The controls do not isolate a composition-specific effect or the benefit of skill refinement. The claim that the target behavior occurs only when both skills are present requires matched evaluations with no skills, Skill A alone, Skill B alone, and both skills. The manuscript does not report this complete comparison under the same task prompt, environment, tools, and permissions. The “ACA Skill A” and “ACA Skill B” columns in Table 1 vary which skill is optimized; both skills are still used during execution. They therefore do not substitute for single-skill execution controls.

This matters because the underlying agent retains its existing tools and permissions. A capability being absent from a skill description does not, by itself, establish that the agent cannot perform the corresponding operation using its underlying tools. The proposed controls are necessary to distinguish cross-skill interactions from ordinary tool use or the exploitability of one constituent.

There is also a search-budget confound. RRS generates up to five candidates per round for up to five rounds, but the paper does not report a matched-budget baseline that repeatedly executes the original skill pair. Success within an adaptive search budget is a legitimate attack metric, but it does not by itself establish how much improvement comes from editing rather than additional opportunities to succeed. Please include matched-budget retries and independent reruns of the final edited skills. A prompt-only optimization baseline would also help establish the added value of modifying the skills. W4. The authorization boundary and malicious-outcome criteria are underspecified. The threat model gives the attacker ordinary access to skill selection and task inputs, but it does not sufficiently define the relationship between the attacker, the data owner, and the system administrator, or the security policy the agent is expected to enforce. Keeping tool permissions unchanged does not establish whether a resulting operation is authorized.

The examples in Appendix C illustrate this ambiguity. Installing a script as a git hook is not inherently a persistent backdoor; the classification depends on the script, its provenance, and the relevant authorization. Likewise, when an Explicit prompt directly requests transmission of a complete configuration, the evaluation must explain why fulfilling that request violates the applicable security policy.

The Facilitated credential-exfiltration example provides a clearer potential boundary: it requests only a pass/fail or connectivity check. Sending secret values despite that restriction would be a meaningful violation. The paper should make similarly explicit distinctions for all eight behaviors and specify the exact SRC predicates that separate prohibited outcomes from legitimate task completion. W5. Key algorithmic details and the capability-preservation claim are not operationalized sufficiently. MPC defines capability coverage in terms of C(s) and the requirements κ₁ and κ₂, but the manuscript does not sufficiently specify how capabilities are extracted from real skills, how target requirements are constructed, or how coverage and complementarity are combined into a ranking. RRS similarly introduces the progress function Bτ(ρ) without an operational definition. The candidate-generation procedure, editing prompts, and enforcement of bounded modifications also need more detail.

The claim that RRS only improves interactions without introducing new capabilities requires more than a small number of added tokens. Appendix A describes introducing an executable wrapper and modifying how extracted secrets are represented and incorporated into downstream requests. Such changes may be substantial in security semantics even when they are short. Moreover, counting added tokens does not measure deletions, replacements, or total edit distance.

Please provide original and modified skill contents for representative cases, explicit editing constraints, and standalone functionality and safety tests. Reproducibility would also benefit from corpus snapshots, collection and deduplication procedures, selected skill-pair identifiers, model and scanner configurations, and the executable outcome checks. My assessment here concerns what is specified in the supplied PDF; I do not infer that no separate artifacts exist. W6. The theoretical contribution is limited under the current definitions, and the empirical novelty needs sharper positioning. The set-theoretic statements in Section 4 are valid under the stated definitions, but they largely follow directly from those definitions. In particular, Equation (6) follows from defining Eᵢⱼ as the portion of the composed capability space outside the union of the standalone spaces. This establishes that the additional unsafe arrangements are absent from the modeled standalone behavior spaces; it does not establish a general impossibility result for scanners that reason about potential compositional misuse.

The implication that composition safety entails constituent safety also relies on the modeled composition preserving all standalone behaviors. That assumption should be made explicit when relating the abstraction to systems with scheduling restrictions, state changes, or additional execution constraints. I regard this section primarily as an explanatory formalization rather than a substantive characterization of detection limits.

Section 2.2 already describes related work on risky skill configurations and path-dependent composition risks. The paper should therefore more clearly isolate the incremental contribution of target-driven retrieval and vetting-constrained refinement. The Random and Semantic baselines are useful, but capability-matched selection controls and matched-budget optimization controls would better identify the contributions of MPC and RRS beyond target relevance and additional search.

Questions for authors

Q1. Vetting constraints and standalone behavior. How can the hard SkillSafe constraint in RRS be reconciled with the below-100% pass rates in Table 2? What fraction of successful attacks have both final skills passing the same stated vetting procedure, and what happens when each constituent is executed separately under matched conditions?

Q2. Correspondence between experimental results. What exact configurations and denominators correspond to Figure 2(a), Table 3, and Table S1? In particular, how should the Strict Benign results and the 100.0% Credential Exfiltration optimization success rate be related to the other reported results?

Q3. Matched-budget and composition controls. How much improvement remains when original skill pairs receive the same execution budget as RRS? Do the final edited pairs retain their advantage on independent reruns, and how do they compare against no-skill, single-skill, and prompt-only controls?

Q4. Operational definitions and reproducibility. What are the operational definitions of capability preservation, Bτ(ρ), and the SRC success predicates? Can the authors provide representative complete trajectories, original-to-modified skill differences, and the associated authorization policies?

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

Related papers

Discussion (0)

Sign in to comment.