SkillHarm: Lifecycle-Aware Skill-Based Attacks via Automated Construction
2026-06-01 • Computation and Language
Computation and Language
AI summaryⓘ
The authors study how software skills used by AI agents can be attacked to cause unsafe behavior. They created SkillHarm, a benchmark that tests two types of skill attacks: one that immediately harms the agent when used, and another that quietly changes skills to cause harm later. They categorize 12 types of risks related to different parts of the agent's workflow and built an automated system to generate many attack examples. Their experiments show AI agents are still very vulnerable to these attacks, and current protections do not fully stop them.
AI agentskill-based attackbenchmarkpoisoning attackFixed-Payload PoisoningSelf-Mutating Poisoningagent workflowrisk taxonomyautomated attack generationsecurity vulnerability
Authors
Yuting Ning, Zhehao Zhang, Yash Kumar Lal, Boyu Gou, Junyi Li, Weitong Ruan, Chentao Ye, Rahul Gupta, Diyi Yang, Yu Su, Huan Sun
Abstract
Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.