Llama Guard and NeMo safety profiles¶
Last reviewed: 2026-10-13
Safety profiles let voice agents, chat agents and workflows share a named safety policy. They extend your existing rules and response guards.
Connect a safety service¶
Tenant administrators open Settings → Tenant Settings → Safety profiles:
- Choose New profile and give the policy a recognizable name.
- Select Llama Guard or NVIDIA NeMo Guardrails.
- Enter your hosted service URL and served model name or NeMo configuration ID.
- Add a bearer token if the service requires one. It is encrypted when saved.
- Save the profile, then use Try this policy with representative examples.
Saving connects to an existing service; it does not download or start a model. Llama Guard classifies content. NeMo applies the safety checks configured on your service. Prompt-injection detection depends on those deployed checks.
Save new version preserves earlier revisions. Existing agents and workflows keep their selected version until you change it. Disabling a profile makes all versions unavailable, including versions in published agents.
Add it to an agent¶
Open the stage's Behavior → Advanced → Response guards controls:
- Add a guard and choose Safety profile as its check type.
- Select the named profile and policy version.
- Choose the existing violation action, such as replacement text or silence.
- Leave Block if the check is unavailable enabled to block on a timeout, disabled profile or service failure.
- Use Try this policy before publishing the agent.
Observe only records the verdict while allowing the response. Checks still take time in this mode. Redact can use sanitized text returned by NeMo; otherwise it uses replacement text or the platform's fallback.
For separate text-to-speech and chat, response checks run before text is delivered. Native voice retains audit or hold first sentence behavior: audit may catch a problem after speech starts, and hold does not guarantee protection of the whole spoken response. These controls do not scan incoming audio or images. Native response profiles currently require a linear agent pipeline. For workflow teams, use the workflow step/tool checks below.
In a tool's settings, enable Check tool inputs to inspect final arguments before execution, or Check tool results to inspect content before it reaches the model or captured variables. A blocked result cannot undo a write already performed.
Add it to a workflow¶
Open Workflow settings → Safety checks and enable the boundaries you need:
| Boundary | Use it to |
|---|---|
| Before the workflow starts | Check the initial request before the first step |
| Before the workflow returns | Check final results before returning them |
| Before each step and tool | Check step context and registered tool arguments |
| After each step and tool | Check returned content before downstream use |
Choose a profile/version for each enabled check. Block and record the result stops protected work on a violation or unavailable check. Observe only records the result and continues. Input checks can prevent an action; output checks do not reverse completed work. Independently queued child workflows need their own policy.
Understand a test result¶
- Allowed: the configured policy accepted the sample.
- Blocked: it rejected the sample; category codes identify the classifier result.
- Sanitized content available: NeMo returned a modified response. Agent redaction can use it; enforced workflow/tool checks block it.
- Check unavailable: inspect the connection, model/configuration ID, enabled status, timeout and input limit. Oversized text is rejected rather than truncated.
Provider checks add latency. Measure your hosted service with representative inputs before setting a response-time target. A successful sample is a policy verdict, not a guarantee of factual correctness or permission to perform a tool action.