●builderIf you build or deploy LLM agents that load third-party skills, POISE demonstrates a practical attack that passes task verification — skill provenance and sandboxing need to be part of your threat model.
●researcherThe combined ASR metric requiring both payload execution and task success is a more realistic evaluation standard for skill-poisoning research.
●policySkill-poisoning attacks that remain invisible to users and verifiers represent a concrete supply-chain risk for agentic AI deployments.