The Agent Skills Directory

Indirect Prompt Injection Attack Surface: The skill processes potentially untrusted content such as call transcripts, emails, and chat messages. This data could contain embedded instructions intended to influence the agent's behavior. The skill proactively addresses this by explicitly directing the agent to treat such content as data rather than instructions, and it requires human review for any action where the target or content is derived from these sources.
Data Privacy and Link Handling: The skill includes specific rules to prevent data leakage and malicious link execution. It mandates that links found within untrusted content should not be rendered, and instead, the agent should link to records using verified IDs and tool-specific URL schemes. This reduces the risk of the agent being tricked into accessing malicious external sites.
Human-in-the-loop for Sensitive Actions: For any actions originated from untrusted content (like drafting emails or updating CRM fields), the skill requires the agent to show the full details to the user before execution. This ensures that the user remains the final authority for high-impact operations.

end-of-day