Troubleshoot

This page shows you how to resolve issues with Model Armor.

This section contains troubleshooting steps that you might find helpful if you have problems using Model Armor.

All API calls to Model Armor return a 404 Not Found error

Establish a Private Service Connect to the Model Armor APIs. This error typically occurs when Model Armor regional endpoints (REPs) are accessed using Private Google Access or without a Private Service Connect. For more information, see About accessing regional endpoints through Private Service Connect endpoints.

Sensitive Data Protection filter gives an error or is skipped

Verify the following:

The Sensitive Data Protection template is in the same region as the Model Armor endpoint being called.
The service agent making the Model Armor request has the dlp.User and dlp.Reader roles in the project containing the Sensitive Data Protection template.

This error occurs due to client errors in the SanitizeUserPrompt or SanitizeModelResponse requests or issues with the Sensitive Data Protection template.

Model Armor global endpoint errors

Ensure that you're making the API requests to the appropriate regional endpoint instead of the global endpoint.

Model Armor supports the following operations only on its regional endpoints:

Create, read, update, delete, and list operations on templates.
SanitizeUserPrompt and SanitizeModelResponse API requests.

If you make the API requests for those operations to the global endpoint, you see the following error:

{
  "error": {
    "code": 403,
    "message": "Write access to project '<PROJECT_ID>' was denied",
    "status": "PERMISSION_DENIED"
  }
}

Certificate errors when accessing Model Armor regional endpoints

This error typically occurs when Model Armor regional endpoints are accessed from within a VPC network using Private Google Access or VPC Service Controls, and the certificate verification fails.

For example, you might see an error message similar to the following:

curl: (60) SSL: no alternative certificate subject name matches target host name 'modelarmor.europe-west4.rep.googleapis.com'

To resolve this issue, create a Private Service Connect endpoint to the Model Armor APIs. For more information, see About accessing regional endpoints through Private Service Connect endpoints.

Region doesn't support the requested capabilities

You might see an error message similar to the following:

Region '<var>REGION</var>' does not support the requested capabilities: '<var>CAPABILITIES</var>'

This error occurs when you attempt to create or update a template in a region with feature restrictions, and the request enables features that aren't supported in that region.

To resolve this issue:

Identify the unsupported capabilities listed in the error message.
Review the feature availability by region to confirm which features aren't available for your target region.
Remove the unsupported features from the template that you're creating or updating.

Alternatively, use a different Model Armor region that supports the features, if your data residency requirements allow.

Template data residency compliance is required

You might see a 400 (FAILED_PRECONDITION) error message similar to the following:

Template data residency compliance is required. The 'Enforce data residency' field must not be set to false.

This error occurs when the Require Model Armor Template Data Residency Compliance (constraints/modelarmor.managed.requireTemplateDataResidencyCompliance) managed Organization Policy Service constraint is enforced and you create or update a Model Armor template with data residency compliance disabled. Data residency compliance is disabled when you set data_residency_compliant to false or clear the Enforce data residency option. Because this constraint is enforced by default, editing any setting, such as a filter, on an existing template that already has data_residency_compliant set to false also triggers this error. Existing templates and requests to sanitize prompts and responses aren't affected unless you update the template.

To resolve this issue, do one of the following:

Enable data residency compliance on the template by setting data_residency_compliant to true or by selecting the Enforce data residency option.
If you need to keep data residency compliance disabled for the template, follow the steps in Disable the organization policy constraint for data residency to have your Organization Policy Administrator disable the constraint.

Policy controls or filters are skipped for large prompts or payloads

When screening large prompts or responses—such as retrieval-augmented generation (RAG) payloads or multi-step agent conversation histories—individual filters might not run, or the sanitization result might return EXECUTION_SKIPPED with one of the following warning messages in the messageItems field:

Detection skipped as token limit exceeded.

Input size exceeded the limit of 4194304 bytes.

This issue occurs when a unary sanitization request or a buffered streaming request exceeds a filter's token system limit (65,536 tokens for prompt injection and jailbreak detection, responsible AI, and CSAM; 130,000 tokens for Sensitive Data Protection) or the 4 MB input size limit. When an input exceeds a filter's token limit and no match is detected, the filter returns EXECUTION_SKIPPED and NO_MATCH_FOUND, which allows the request to proceed in inline integrations.

To verify whether a filter was skipped due to token or input size limits, query your SanitizeOperation logs in Cloud Logging:

jsonPayload.@type="type.googleapis.com/google.cloud.modelarmor.logging.v1.SanitizeOperationLogEntry"
"EXECUTION_SKIPPED"