Exam-style
A customer complains that a Bedrock-backed assistant promised them a refund the company does not offer, and quotes the reply. The team believes the assistant could not have produced that text. AWS CloudTrail is on, Amazon CloudWatch metrics and alarms are in place, and the application writes its own logs. Which record reconstructs the actual exchange?
Reveal the answer
B. Amazon Bedrock model invocation logging, which captures the prompt sent and the completion returned and delivers them to a destination the team owns
Only Amazon Bedrock model invocation logging stores the text. It captures the request and the response, including the prompt and the completion, and delivers them to an Amazon S3 bucket or an Amazon CloudWatch Logs group the team owns. It is opt-in and off by default, which is why teams discover they need it on the day somebody asks what was said. AWS CloudTrail records the API call rather than its content: which principal called InvokeModel, from which account, at what time, against which model. That is enough to prove a call happened and to find who made it, and it will never show a sentence the model produced. Amazon CloudWatch carries counts and shapes: invocations, latency, error rates, guardrail intervention counts, all of which are the raw material for alarms and for threat detection, and none of which is a transcript. The application logs hold the request that reached the service, which is the part the customer typed plus whatever context the code attached; the completion came back from Bedrock and only appears there if somebody wrote code to log it, and logging generated text by hand is how it ends up in the wrong store. Replaying the question proves nothing either, since a foundation model is not deterministic and the retrieved context has moved on. The consequence worth carrying: once invocation logging is on, that bucket or log group holds every prompt and every answer, which makes it the most sensitive store in the workload and gives it its own encryption key, its own access policy and its own retention setting.
Q. A customer quotes a reply the assistant supposedly gave. CloudTrail, CloudWatch and application logs are all running. Which record holds the exchange?
A. Amazon Bedrock model invocation logging, which captures the prompt and the completion and writes them to a bucket or log group you own. It is off until somebody switches it on.
Why? Sort the four logs by what each one holds. AWS CloudTrail holds the call: principal, time, region, model, no text. Amazon CloudWatch holds the numbers: invocations, latency, errors, guardrail interventions, the material for alarms and for threat detection. Application logs hold the request that arrived at your service, which is application security territory and not the model’s answer. Invocation logging holds the words, and it is the only one that does. That is what audit trail and logging requirements for AI interactions mean in practice: two records at minimum, one for who called and one for what was said. Turning the second one on moves the sensitive data problem rather than solving it, so give that destination a customer-managed key, a tight access policy and a retention rule before the logs start arriving.