Amazon Bedrock

Bedrock is zero-retention and no-training by default, but named model routes have current abuse-review retention exceptions.

Plan Scope

Amazon Bedrock model inference; model-specific abuse rules and customer-enabled stateful features can differ from the default.

Prompt Retention

Bedrock defaults to zero data retention and does not store model inputs or outputs. Current exceptions include up to 30-day retention for classifier-flagged OpenAI GPT-5.4/5.5/5.6 traffic and all Anthropic Claude Fable 5 traffic, plus legally required CSAM handling.

Response Retention

Same default and model-specific exceptions as inputs. Customer-created agents, knowledge bases, invocation logging, batch files, and other AWS storage persist under their configured lifecycles.

Ordinary Logging

Bedrock retains metering and operational metadata, and customers can deliberately enable model-invocation logging to CloudWatch or S3. Content is automatically screened for abuse.

Model Training

AWS says it does not use Bedrock inputs or outputs to train or improve base models and does not share them with third-party model providers for that purpose.

Product Improvement

Customer content is excluded from base-model improvement; aggregate service telemetry and customer-provided feedback are separately governed.

Human Or Operator Access

Bedrock uses a zero-operator-access model by default. Stored abuse exceptions can be reviewed only for the specified safety purpose; customer support or customer-configured logs create separate access paths.

Subprocessors And Routing

AWS hosts access to Amazon and third-party foundation models without sharing prompts with the original model providers, but model licenses and use policies still apply.

Deletion Controls

Ordinary stateless payloads are not stored. Customers control CloudWatch, S3, agents, knowledge bases, batch inputs, and other feature state through AWS retention and deletion settings.

Caveat

Model-specific safety exceptions now prevent treating every Bedrock route as strict ZDR. Inspect the selected model and any enabled logging or stateful feature.