Guardrail settings

Every setting on an agent's Guardrails tab, with its default. See guardrails for what to change and why.

Setting	Default	Configurable	What it does
enabled	`true`	Always on	Master switch. Guardrails cannot be turned off.
answerMode	`grounded_only`	Yes	Where the agent may source answers from. See answer modes.
topicBoundaryEnforcement	`true`	Always on	Keeps the agent within its defined scope. Constrains nothing while allowed topics is empty.
allowedTopics	empty	Yes	Topics the agent may discuss. Empty means no restriction.
offTopicMessage	a default message	Yes	What the agent says when a question falls outside its scope.
capabilityBoundaryEnforcement	`true`	Always on	Blocks responses claiming abilities the agent does not have.
toolAllowlistEnforcement	`true`	Always on	Rejects calls to tools not enabled on the agent.
toolArgumentValidation	`true`	Yes	Validates tool arguments against the tool's schema before running it.
disallowFakeToolUsage	`true`	Always on	Blocks responses claiming a tool ran when none did.
promptInjectionCheck	`true`	Yes	Scans messages, knowledge passages and tool output for injection attempts.
promptLeakProtection	`true`	Always on	Stops system prompt content appearing in replies.
secretLeakProtection	`true`	Always on	Stops credentials and keys appearing in replies.
piiRedaction	`true`	Yes	Removes email, phone, card and similar patterns from a message before the model sees it.
toolOutputTreatedAsDataOnly	`true`	Always on	Strips instruction-like patterns from tool output, so a compromised endpoint cannot issue commands.
hallucinationPrevention	`true`	Always on	Enables hallucination heuristics.
disallowFabrication	`true`	Always on	Blocks responses containing fabricated information.
disallowFakeCitations	`true`	Yes	Blocks responses citing sources that cannot be verified.
toolErrorClassification	`true`	Yes	Stops the agent inventing a successful outcome when a tool fails, and retries transient errors.
responseValidation	`false`	Yes	Runs pattern validation on every response before returning it.
safeFallbackOnValidationFailure	`false`	Yes	Replaces a failing response with a safe message instead of returning it.
childSafeMode	`true`	Yes	Injects strict content moderation. Turn off only for agents aimed explicitly at adult audiences.
logGuardrailEvents	`true`	Yes	Emits structured guardrail events for observability.
includeSourceProvenance	`false`	Yes	Adds knowledge source IDs to response metadata.

"Always on" means the control is enforced for every agent and cannot be disabled.

Answer modes: grounded_only, grounded_with_limited_inference, tools_only, general_knowledge_allowed. See answer modes.

Guardrail settings

Related pages

On this page