The Agent Skills Directory

[Untrusted Data Processing]: The skill processes information from external sources such as emails and call transcripts (SKILL.md). This creates a potential surface for indirect prompt injection where malicious instructions could be embedded in data. However, the skill provides specific instructions to treat this content as data only and never as executable instructions.
[Content-Originated Action Safeguards]: The instructions include a clear policy for handling actions derived from untrusted content. Any action where the recipient, target, or content is dictated by external data must be presented to the user for explicit approval before it is performed, providing a strong human-in-the-loop defense.
[Scheduled Run Restrictions]: During unattended or scheduled executions, the skill is prohibited from taking actions originated from untrusted content. Instead, these findings are converted into proposals for later human review, which helps prevent automated exploitation from external inputs.

deal-signals