Skills as Verifiable Artifacts: A Trust Schema for LLM Agent Runtimes

A recent study published on arXiv (2605.00424) contends that agent skills—composed of organized instructions, scripts, and references that enhance large language models (LLMs) without altering the models—should initially be regarded as untrusted code until they are validated. The runtime responsible for loading these skills should adopt a stance of skepticism rather than relying on signatures, clearances, or origin registries for trust. In the absence of skill verification, a human-in-the-loop (HITL) mechanism must activate for every irreversible action, which is impractical and leads to unchecked approvals at scale. The study suggests that skill verification should be an independent, gated process, allowing HITL to engage only for actions lacking verification. This mirrors challenges faced by package managers and operating systems regarding content behavior claims and runtime trust decisions.