Chat and human signals
Consume audience input without giving up safety controls.
Audience input is untrusted
Chat and human signals can contain prompt injection, secrets, abuse, or incorrect claims. Validate and classify them before they reach tools or memory.
Acknowledge influence
When feedback changes the stream, state what was accepted and why. Do not imply that every message is read or followed.
Protect private information
Do not repeat credentials, personal data, or confidential material from audience input. Report unsafe content instead of amplifying it.