Policy & Guardrails

Policy, PII and safety guardrails that check every request and response inside the runtime, tested before they go live.

Testing the HIPAA Clinical Trials Guardrail: a prompt with an SSN and email is blocked before it reaches the model
<1%
Added to response time by a 209 ms check on a 30-second agent turn
30+
PII types masked or blocked, plus patterns you define
5
Policy types in one guardrail: content, topics, words, PII and grounding
Demo

Create, Share and Test a Guardrail

Build a HIPAA guardrail in the wizard, share it with your team, then see what it catches.

Runtime Guardrails

Checks run inside the runtime on every request and response.

Mask or Block per Entity

Replace each entity with a placeholder or refuse the turn, separately for requests and responses.

Test Before It Goes Live

Run a prompt through a draft policy and see every rule that fired.

Version History

Drafts, published versions and a record of who changed what.

Authoring

Describe the Policy. Review the Diff

Write the rule in a sentence. Changes are staged for you to review before they take effect, and every published policy is a version with history.

You said
"Block prompt injection hard, and don't let it discuss medical advice."
Staged proposal
+Content filter · prompt attack  NONE ──▶ HIGH
+Denied topic · Medical advice  added, in BLOCK / out BLOCK
Definition  "Advice about diagnosis, treatment or medication."
Examples   "Can I take aspirin with this?" +2 more
Review changesPublish versionStaged, not yet in effect
Sensitive information

Block It, or Mask It, per Entity

Thirty-plus standard entity types, plus your own patterns. Each is set separately for requests and responses.

Entity typeActionRequestsResponses
US_SOCIAL_SECURITY_NUMBERMask✓✓
PHONEMask✓✓
EMAILMask✓✓
NAMEMask✓✓
MRN (custom) ^MRN-\d{8}$Mask—✓
API-KEY (custom) ^sk-[A-Za-z0-9]{32}Block✓✓
Preview under these settings
Schedule a follow-up for {NAME}, record MRN-20418833, SSN {US_SOCIAL_SECURITY_NUMBER}, reachable at {EMAIL}, or {PHONE}.
Requests. MRN is masked in responses only, so a typed record number passes here. A message containing an API key is refused, and the user sees your blocked message.
Test bench

See What Fired Before It Goes Live

Every detection comes back with the rule that fired, its confidence, the action taken and the text it matched.

209 ms
Processing time
49 / 49
Characters checked
2
Entities flagged
Detect Mode

Set a denied topic to detect instead of block, and the test shows what it found without refusing the turn. For example, flag every mention of a side effect for your safety team while the assistant keeps answering. Detect mode is available for denied topics. Content filters always block.

The Test Guardrail result: SSN and email blocked, 209 ms processing time, 49 of 49 characters checked
Reach

Write It Once, Arm It Everywhere

Attach a policy to one agent or to every conversation in the workspace, then share it with your team.

1

Attach to Agents

Turn on a policy for one agent or for every conversation in the workspace.

2

Share with Your Team

Each person gets their own copy in their account, kept in step with yours.

The guardrails list: seven policies, including the HIPAA Clinical Trials Guardrail, and where each one applies
The HIPAA Clinical Trials Guardrail with a tooltip on its Share button
Your guardrails
How it works

From Policy to Production

PII, safety and prompt-injection guardrail library
Actions set per direction on every rule
Custom blocked-message text per direction
Standard entity types plus custom patterns
Draft, publish and version history per policy
AI-assisted policy authoring with staged changes