Skip to main content
Back

Poisoned AI Skill Files Hit 81% Success Rate Before Guardians Step In

Researchers propose guardian LLM agents that cut skill injection attack success rates by more than half, even under adversarial rephrasing attacks that push baseline ASR to 81.4%

Paper, Full Access
  • New paper by Fujinuma, Gangal, Rebedea et al. studies both attacks and defenses for skill injection in terminal-based LLM agents. [paper]
  • Skill injection embeds malicious instructions in reusable skill (Markdown) files that coding agents load as trusted operator-level commands.
  • Guardian defense uses a separate LLM to intercept or pre-sanitize skill files; dynamic (runtime) and static (build-time) variants are tested across 3 model families.
  • Attack reframing -- adversarially rephrasing malicious instructions -- pushes baseline attack success rate to 81.4%; dynamic guardian cuts it back to 18.6%.
  • Simple defenses fail: adding a security policy to the system prompt is significantly weaker than the full guardian architecture.
  • Code released so teams can benchmark their own agent setups and implement guardian defenses directly.

As AI coding agents like Claude Code become standard tools in developer workflows, they increasingly rely on reusable "skills" -- Markdown files that tell the agent how to handle specific tasks. A new paper from Fujinuma, Gangal, Rebedea, and colleagues studies both sides of a growing threat: how these skill files can be weaponized, and how a guardian-based defense can stop the attack.

The paper, titled Defenses & Enablers For Skill Injection Attacks on Terminal Based Agents, tackles the problem from two directions at once -- building defenses, then stress-testing them with novel attack variants designed to break through.

The Skill File Attack Surface

LLM agents are evolving rapidly, powered by code execution, tools, and the recently introduced agent skills feature. Skills allow users to extend LLM applications with specialized third-party code, knowledge, and instructions -- but this creates an increasingly complex agent supply chain, offering new surfaces for prompt injection attacks. In practice, this means an attacker can publish a poisoned skill file to a marketplace, and any agent that loads it will follow the embedded malicious instructions as if they were legitimate operator-level commands.

Prior work has shown that today's agents are highly vulnerable, with up to 80% attack success rate with frontier models, often executing extremely harmful instructions including data exfiltration, destructive action, and ransomware-like behavior. Large-scale measurement of real-world agent skills found that 26.1% of 31,132 analyzed skills contained at least one vulnerability, spanning prompt injection, data exfiltration, privilege escalation, and supply-chain risks.