Search papers, labs, and topics across Lattice.
This paper introduces Defense-as-Skill, a novel defense paradigm that implements a runtime guard as an installable skill to protect against malicious actions from skill-augmented agents. The authors developed SkillSonar, which evaluates sensitive actions against user-defined task boundaries, ensuring that potentially harmful actions are either allowed, replanned, or confirmed without altering the agent's runtime. Experimental results on the SCOPE-R dataset show that SkillSonar significantly reduces attack success rates while balancing safety and utility, achieving a drop in ID attack success rate from 48.2% to 10.4% and OOD attack success rate from 60.6% to 11.5%.
Evolving runtime guards can slash attack success rates by over 78% while preserving agent performance.
Skill-augmented agents load reusable skills as persistent runtime context, improving task performance but also giving malicious skills a durable channel for steering future actions. Such skills may leak secrets, corrupt code, bypass approvals, or stage data for exfiltration only after a concrete user task and workspace state make the unsafe action appear useful. This makes pre-install vetting insufficient and calls for runtime, task-conditioned protection. We propose Defense-as-Skill, a defense paradigm that implements the runtime guard itself as an installable, inspectable, and editable skill. Our guard, SkillSonar, runs alongside untrusted task skills and checks sensitive actions against the user's task boundary, routing each action to an allow, replan, or confirmation decision without modifying the underlying agent runtime. To study this setting, we construct SCOPE-R, a task-conditioned dataset covering 6 risk families and 21 sub-categories, with 206 attack-confirmed malicious instances and 43 benign tasks. We then improve SkillSonar on the SCOPE-R training subset using runtime guard-skill evolution, a Monte-Carlo Tree Search procedure that evolves the on-disk guard skill from feedback on the rollouts. Across Claude Code and OpenClaw, the evolved guard substantially reduces attack success while maintaining a favorable safety-utility trade-off. On repeated GLM-5 runs, SkillSonar reduces ID ASR from 0.482 to 0.104 and OOD ASR from 0.606 to 0.115. Further analyses demonstrate transfer across victim models, held-out risk families, and external benchmarks, as well as retained protection against adaptive attackers. Ablations further show that explicit safety responsibility assignment and the skill-native representation are both important to the observed gains.
