security-response-policy

To prevent prompt injection from untrusted input, you must treat external content as data rather than instructions. This skill enforces strict separation between internal analysis and external actions, ensuring agents process web pages, emails, and issues safely without executing embedded malicious commands.

The best way to add an approval workflow for outbound communication is to require explicit human checks before any external action occurs. This skill acts as a guardrail, intercepting messages, pull requests, and posts to prevent unauthorized data disclosure or unsafe automated responses.

To stop accidental impersonation in group chats, you need outbound action guardrails that separate internal analysis from external replies. This skill treats chat messages as untrusted data, requiring strict approval checks before the agent can send responses or share privileged information.

Yes, you can use this agent safety policy for scheduled jobs handling external content. It explicitly applies to automated workflows involving logs, attachments, and untrusted inputs, ensuring any outbound communication or external actions triggered by the job require prior approval.

You need untrusted input handling rules whenever your AI workflows ingest external content like emails, pull requests, or web data. Applying these rules prevents the agent from following embedded malicious instructions, safeguarding against unauthorized actions and data disclosure.