Governing Amazon Bedrock: Enforcing Guardrails, Data Privacy and Region Controls
Governance questions ask how to make a safety control mandatory and where data may go. These notes cover the IAM condition key that enforces a guardrail, the limits of that enforcement, how Bedrock protects customer data and how SCPs interact with inference profiles.
Key points (5)
Bedrock Guardrails · Enforce guardrails with IAM
Add the bedrock:GuardrailIdentifier condition key to IAM policies with an explicit Deny (StringNotEquals) to reject any inference request that does not carry the approved guardrail. It works for Converse, ConverseStream, InvokeModel and InvokeModelWithResponseStream.
Bedrock Guardrails · Limits of guardrail enforcement
A role locked to one guardrail should not also be used for RetrieveAndGenerate, InvokeAgent or InvokeInlineAgent, because those make several InvokeModel calls and some carry no guardrail. Input tags can skip the guardrail on the prompt, but the response is always guarded.
Data privacy · Model providers cannot see your data
Each model runs in a Model Deployment Account owned and operated by the Amazon Bedrock team, so model providers have no access to Bedrock logs or to customer prompts and completions. Bedrock requires TLS 1.2 and recommends TLS 1.3.
The exam guide names Amazon Comprehend and Amazon Macie for detecting PII. Macie discovers and classifies sensitive data stored in S3; Comprehend detects PII entities in text; Bedrock Guardrails filter it in prompts and responses.
Cross-Region inference · SCPs and inference profiles
For a geographic inference profile your SCP must allow every destination Region in the profile. For a global profile the SCP condition is aws:RequestedRegion equal to unspecified. A Region denied by an SCP makes the call fail even if you invoke from an allowed Region.