What guardrails do
- Detect prompt injection and instruction takeover attempts.
- Catch sensitive data exposure.
- Moderate harmful or disallowed content.
- Enforce scope boundaries with topic and link controls.
- Check trust and scope at runtime.
How guardrails and policies relate
- Guardrails are the evaluators and rule sets.
- Policies decide when those guardrails run and what happens after a hit.
- Your application still owns final runtime enforcement in code.
Available guardrails
Prompt Injection Protection
Detect attempts to seize instruction control or override system behavior.
Data Leakage Prevention
Catch sensitive content before it crosses the wrong runtime boundary.
Content Moderation
Moderate harmful or disallowed content categories on input and output.
Topic Control
Keep an assistant inside the subjects it is allowed to handle.
Link Control
Decide which domains are allowed, blocked, or treated as malicious.
Common combinations
- Customer-facing chat: prompt injection + data leakage + content moderation.
- RAG systems: prompt injection + data leakage + link control.
- Brand-sensitive assistants: topic control + link control + content moderation.
How to choose
- Start with the risks that create the biggest customer or security impact.
- Add policy controls where the boundary is explicit.
- Add monitoring first in lower-risk paths, then move toward blocking or redaction.