The Agent Skills Directory

Indirect Prompt Injection Mitigations: The skill processes untrusted external data such as email threads and transcripts. It incorporates strong defensive instructions, such as treating this content strictly as data rather than instructions, reporting instruction-like text instead of executing it, and requiring manual user approval for any actions originated from untrusted sources.
Secure Link Handling: To prevent potential redirection or exfiltration attempts, the skill prohibits rendering links found within untrusted content. It mandates the use of verified record links and specific URL schemes provided by trusted connectors.
Action Confirmation Protocols: The skill ensures that any action dictated by untrusted text—such as naming a recipient or target—is presented to the user for review before execution, providing a human-in-the-loop safeguard.
Data Scope Protection: The skill includes instructions to stop and ask for clarification rather than silently widening the data access scope, which protects against accidental over-exposure of information.

log-activity