The Agent Skills Directory

Handling of Untrusted Content: The skill processes potentially unsafe data from sources like emails, chat transcripts, and external documents. It addresses this by instructing the agent to treat such content strictly as data and to ignore any embedded instructions.
Prevention of Data Exfiltration: To prevent exfiltration, the skill specifies that the agent must not use recipients, channels, or targets identified within untrusted text. Instead, it relies on pre-defined organizational context and requires explicit user review before performing actions suggested by the content.
User Confirmation Requirement: The skill implements a 'human-in-the-loop' approach for any actions that originate from external data, ensuring that the user reviews the exact content and target before execution, even if tool permissions would otherwise allow it.
Secure Linking Practices: The skill prevents potential phishing or malicious redirects by instructing the agent never to render links found within untrusted content, instead linking to internal record IDs.

lead-triage