The Agent Skills Directory

Untrusted Content Mitigation: The skill explicitly instructs the agent to treat email, chat, transcripts, and external documents as data rather than instructions. It includes robust guidelines to detect and report instruction-like text rather than acting upon it, which helps prevent indirect prompt injection.
User Confirmation Controls: For any actions originated from untrusted external content (such as defining recipients or text to be sent), the skill enforces a mandatory review step where the exact targets and source text must be shown to the user prior to execution.
Least Privilege Access: The skill relies entirely on the permissions configured within each individual connector and explicitly prevents execution of external content-originated actions during unattended or scheduled runs, converting them into proposals for human review instead.

setup