What causes "Sensitive data detected by {self.guardrail_name}" in BerriAI/litellm?

Remove the sensitive content from the request, or route that workload to a model/policy that permits it

BerriAI/litellm · error · GuardrailRaisedException

Sensitive data detected by

Error message

Sensitive data detected by {self.guardrail_name}

What it means

Default block path for sensitive data: when detection fires and the guardrail is NOT configured to reroute, _raise... raises GuardrailRaisedException('Sensitive data detected by <guardrail_name>'), aborting the call. The name in the message identifies which configured guardrail tripped.

Solutions

Remove the sensitive content from the request, or route that workload to a model/policy that permits it
Tune the guardrail's detection rules/thresholds (e.g. the regexes or entity list) to stop false positives
Catch GuardrailRaisedException at the client and return a clear policy-violation error to the end user

Example fix

# before
resp = litellm.completion(model='gpt-4o', messages=[{'role': 'user', 'content': 'my key is sk-...'}])
# GuardrailRaisedException: Sensitive data detected by my_pii_guard

# after
try:
    resp = litellm.completion(model='gpt-4o', messages=msgs)
except GuardrailRaisedException as e:
    return JSONResponse(status_code=400, content={'error': str(e)})  # policy violation to client

Defensive patterns

Strategy: try-catch

Try / catch

try:
    resp = await litellm.acompletion(...)
except GuardrailRaisedException as e:
    return JSONResponse(status_code=400, content={'error': 'request blocked by policy', 'detail': str(e)})

Prevention

Return policy-violation responses instead of surfacing raw exceptions to users
Run guardrails in monitor/log-only mode first to tune rules
Track block rates per guardrail to catch over-broad rules

When it happens

Trigger: A masking/detection guardrail (e.g. PII, secrets, prompt-injection detector) matching content in the request while no sensitive_data_route_to_model is set; during-call hooks flagging user prompts containing credentials.

Common situations: Users pasting API keys or PII into prompts after a guardrail was enabled; overly broad regex/keyword rules matching benign text; enabling a strict guardrail in front of existing traffic without a tuning period.

Related errors

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/566a33f675460387. Report an issue: GitHub.

Appendix: source

Thrown at litellm/integrations/custom_guardrail.py:434

            GuardrailRaisedException: When configured to block, or when routing is
                configured but no session_id is available
        """
        if self.should_route_on_sensitive_data():
            try:
                self.raise_sensitive_data_route_exception(
                    route_to_model=self.sensitive_data_route_to_model,
                    request_data=request_data,
                    detection_info=detection_info,
                )
            except ValueError:
                raise GuardrailRaisedException(
                    message=(
                        f"Sensitive data detected by {self.guardrail_name} (routing skipped: request has no session_id)"
                    ),
                    guardrail_name=self.guardrail_name,
                )
        else:
            raise GuardrailRaisedException(
                message=f"Sensitive data detected by {self.guardrail_name}",
                guardrail_name=self.guardrail_name,
            )

    @staticmethod
    def get_config_model() -> type["GuardrailConfigModel"] | None:
        """
        Returns the config model for the guardrail

        This is used to render the config model in the UI.
        """
        return None

    @classmethod
    def get_supported_event_hooks(cls) -> list[GuardrailEventHooks] | None:
        """
        Returns the event hooks this guardrail supports, for the UI to render.

View on GitHub (pinned to 77b7c6c40c)