Enforcing Guardrails in Amazon Bedrock using IAM

Organizations using Amazon Bedrock often need to ensure all generative AI requests include appropriate content guardrails. Rather than depending on developers to consistently implement these safeguards, AWS Identity and Access Management (IAM) policies can be configured to automatically reject any Bedrock requests that don't include guardrails. This approach creates a systematic enforcement mechanism at the permission level, ensuring unfiltered AI content cannot be generated within your environment. This post demonstrates how to implement IAM policies that require guardrails for all Bedrock model invocations. By establishing these policies, organizations can implement consistent safety standards across all AI applications without relying on individual compliance.

This following AWS IAM policy grants permission to invoke Amazon Bedrock foundation models through InvokeModel and InvokeModelWithResponseStream API actions. However, the policy simultaneously enforces a critical security requirement by explicitly denying these same actions if the request doesn't include a guardrail identifier, effectively making guardrails mandatory for all model invocations.

Another policy is also required to make sure the role being used has permission to apply the guardrail.

If this AWS Lambda is executed with a payload containing a “guardrailIdentifier” parameter which contains a valid Bedrock Guardrail ID the system will produce the expected results. If the “guardrailIdentifier” parameter is missing the system will produce an "AccessDeniedException".

The IAM policies shown here provide a simple yet effective way to enforce Amazon Bedrock guardrails across your organization. Instead of hoping developers remember to implement safety measures, these policies automatically block any AI requests that don't include guardrails. As demonstrated in the Lambda example, properly configured requests work normally, while those missing guardrails are immediately rejected. This approach ensures consistent AI safety standards without requiring individual compliance from every developer.