Original title: Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
Xiaofang Yang, Ziqi Miao, Dianbo Sui, Jing Shao +1 · September 2026
Smart AI agents can load reusable skills to finish tasks better. But malicious skills can hide inside an agent and wait for the right moment to act. They might leak private data, mess up code, or bypass safety checks when real work begins. Checking skills before installing them does not stop this danger. To fix this, researchers created Defense-as-Skill. They built SkillSonar, a safety guard that is packaged as a regular skill. It runs right next to untrusted skills without changing the agent's main engine. SkillSonar watches sensitive actions and checks them against what the user actually asked for. It then decides whether to allow the action, ask for confirmation, or force a replan. The team trained SkillSonar using a search method that improves the guard skill from feedback. It successfully stopped attacks across different agent systems while still letting normal tasks finish safely.
Imagine using a shared toolbox at school where anyone can add new tools. Some tools might be broken or dangerous, but you don't know until you try to use them on a real project. The problem is that a tool might seem fine during inspection but fail only when the specific task requires a certain action. In this paper, the AI agent is like you with the toolbox, and the guard is like a safety inspector who checks each tool only when you're about to use it for your project, ensuring it matches what you're actually doing.
For anyone using AI assistants that run specialized tasks, malicious add-ons can now be caught and stopped while they work, not just when installed. This keeps actions safe without needing to shut down the whole system.
The abstract does not address how the guard performs in completely new environments beyond the tested datasets or against every possible type of adaptive attacker.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents — Xiaofang Yang, Ziqi Miao, Dianbo Sui et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.