c1.4 Log every model invocation with CloudTrail and CloudWatch

Make inference activity auditable

Two different logs answer two different questions:

• AWS CloudTrail records the control-plane and API activity — who called which Bedrock operation, from where, and when. This is on by default for management events and is your audit trail of who did what.
• Model invocation logging captures the data plane — the actual prompts, completions, and invocation metadata for calls made through the bedrock-runtime endpoint. This is off by default and you must enable it explicitly.

Together they let you reconstruct both the administrative actions and the inference itself.

Turn on model invocation logging

Enable invocation logging and send it to CloudWatch Logs (and/or S3). CloudWatch is convenient for metric filters and alarms; S3 is better for cheap long-term retention. The delivery role must let Bedrock write to the destination:

aws bedrock put-model-invocation-logging-configuration \
  --logging-config '{
    "cloudWatchConfig": {
      "logGroupName": "/aws/bedrock/model-invocations",
      "roleArn": "arn:aws:iam::111122223333:role/BedrockInvocationLoggingRole"
    },
    "textDataDeliveryEnabled": true,
    "imageDataDeliveryEnabled": false,
    "embeddingDataDeliveryEnabled": true
  }'

Note: invocation logging only captures calls through the bedrock-runtime endpoint (InvokeModel, Converse, and the OpenAI-compatible APIs on that endpoint). Calls through other endpoints are not captured.

Alarm on the signals that matter

Once logs land in CloudWatch, add metric filters and alarms for things worth waking up for — a spike in guardrail interventions, throttling errors, or an unusual jump in invocation volume from one principal. Prompts and completions can contain sensitive data, so encrypt the destination with your customer managed key (see the KMS control) and scope who can read the log group or bucket.

Keep an eye on cost

• CloudWatch Logs bills for ingestion and storage; prompts/completions can be large, so route high-volume workloads to S3 and keep CloudWatch for a shorter alerting window.
• Set an S3 lifecycle rule to expire or archive old invocation logs.