Review of a documentation-writing AI skill finds risks in trust, permissions, and update handling

Researchers assessed scaffold-docs, an existing AI agent skill that generates documentation, against the OWASP Agentic Skills Top 10 security categories. The skill reads repository content that others may control and writes files, so its safety depends on both its instructions and the agent's permissions. The team reviewed a pinned commit in an isolated workspace without activating the skill, treating embedded instructions as evidence rather than commands. They found no malicious payload, bundled executables, or required remote fetches in that snapshot. The main concern was that ordinary workflow instructions left trust, permissions, and completion state underspecified, which the researchers said mattered more than any hidden command.