Researchers report runtime guard reduces malicious agent-skill attacks

Researchers report in an arXiv preprint that SkillSonar, a runtime guard installed as an agent skill, reduced successful attacks from malicious skills. In repeated GLM-5 runs, attack success rates fell from 0.482 to 0.104 on in-distribution tests and from 0.606 to 0.115 on out-of-distribution tests. SkillSonar checks sensitive actions against the user's task boundary and chooses whether to allow them, replan or request confirmation, without changing the underlying agent runtime. Researchers improved the guard using feedback from task executions. Their SCOPE-R dataset contains 206 attack-confirmed malicious instances and 43 benign tasks across 6 risk families and 21 sub-categories; the reported reductions are experimental results.