Skip to main content

Hallucination protection

Updated on 3 November 2026.

Score model responses for hallucinations and monitor or block them above a confidence threshold. Runs at the response stage.

Configure this control​

Choose the mode used by your gateway. When the Policy Engine is on, published rules govern this control; classic values saved through Edit anyway apply only after the engine is disabled. See What happens to classic settings.

Where it is configured​

Open Govern > Policy Engine > LLM Gateway > Controls > Hallucination Protection. Start from a recipe or a section's add menu, choose Applies to, set protection or a tag rule, then Save to draft. Test the rules before Review & publish. The policy sections and settings are described below.

Sections​

SectionWhat it doesEmpty state
ProtectionSwitch, threshold, action and risk level. Use a recipe or the section add menu.Responses are not checked for hallucinations.
Tag findingsFollow-up rules that tag a severity on the detection result. Use the section add menu.Nothing is tagged on the detection result.

Applies on bedrock, chat, responses and vertex, and only to non-streaming responses. Streaming responses cannot be blocked, so they are not checked.

Settings​

Open a recipe or the add menu in Protection to set:

SettingOptionsDefaultNotes
ProtectionOn, OffOnOff exempts the scope only when its rule outranks a matching On rule.
Threshold0 to 1, step 0.050.8Lower catches more, with more false positives.
When flagged: actionMonitor, BlockMonitorBlock replaces the response with a refusal.
When flagged: risk levelLow, Medium, HighMediumTags the hallucination itself.
SeverityNot set, Very low, Low, Medium, High, Critical, Very criticalNot setApplies to the whole request. Reported only.

Follow-up rules​

Open the add menu in Tag findings to set a follow-up rule. When is one of: a hallucination is detected, the score is at least (0 to 1), or the risk level is (Low, Medium, High). The rule then tags a severity. The highest severity among matching rules is reported.

Scoping and advanced scenarios​

  • Applies to: Everyone, People, Smart group, Application, App tag, Requested model, Provider, API surface, Environment, Tool, Source network, or Except....
  • Threshold: the lowest positive matching threshold applies, whatever its priority. A narrower scope cannot raise that threshold; a value of zero is treated as no threshold.
  • Action and risk level follow the highest-priority matching configuration.
  • On/Off, action and rating use highest priority; On, Block and the higher rating win equal-priority ties. An Off exception must outrank the On rule.
  • Need to key protection off attachment types, streaming or a metadata field? Use Add configuration.

For example, block likely hallucinations for one customer-facing app while the rest of the tenant only monitors, or set a lower (stricter) threshold for one Smart group. Start at Monitor, then move the action or the threshold:

monitor_high_confidence_hallucinationsresponse

runs on responsepriority 500

WhenApplication method typeis any ofchatresponsesbedrockvertex
Then
Hallucination checktrue
Hallucination threshold0.82
Hallucination risk levelhigh
Hallucination actionmonitor