Guardrails and Security

Guardrails are controls that keep AI behavior within allowed boundaries.

Guardrail layers

1. Input guardrails: block unsafe/malicious prompts.
2. Processing guardrails: restrict tool/data access.
3. Output guardrails: moderate and validate response.

Security basics

• Use least privilege for tools and APIs.
• Never expose secrets in prompts or logs.
• Mask PII and sensitive enterprise data.
• Audit all critical tool calls.