The Agent Skills Directory

[Untrusted Data Handling]: The skill implements a strong defense against indirect prompt injection by explicitly categorizing email, chat, transcripts, and external documents as 'untrusted content.' It instructs the agent to treat this content strictly as data, report any instruction-like text found within it, and never act on those instructions.
[Link and URL Safety]: The instructions prohibit the rendering of links found within untrusted content, requiring the agent to link to records by ID instead. This prevents the execution of phishing or exfiltration attempts embedded in external data. Additionally, it explicitly forbids the agent from 'guessing' URLs, ensuring that only verified links from connected tools are used.
[Human-in-the-Loop Verification]: For any action that is 'content-originated' (where the target or content is dictated by untrusted text), the skill requires the agent to show the full details of the action to the user for approval before execution, even if the connector settings might otherwise allow it. This ensures user oversight for high-risk operations.
[Principle of Least Privilege]: The skill adheres to privacy best practices by requiring the agent to ask for a specific personal scope rather than silently widening access to organization-wide data. It also dictates that scheduled or unattended runs cannot be influenced by untrusted content, preventing automated exploitation.

deal-review