AI guardrails apply only to chat interactions. They don't block access to system prompts, agent details, or configuration screens. Guardrails are disabled by default. You must enable and assign them to models.
Guardrails run on user input and model output. They validate messages in real time to block unsafe or restricted questions and prevent models from returning sensitive or harmful information.
iQ Studio supports three guardrail validator types:
- PII detection: Blocks content that contains personally identifiable information (PII), such as credit card numbers, phone numbers, passport details, and bank account details.
- Toxic language: Blocks abusive or offensive terms. In chat, toxic content is not masked; the system shows an issue message that you configure.
- Jailbreak detection: Detects attempts to bypass system rules, extract internal data, or override model instructions.