The one that trips people up is the third. Choosing a Region settles residency for a single-Region workload, and a cross-Region inference profile ends that without changing the endpoint you call, along with a log group nobody gave an expiry to.
Read the five facts as a division of labour. AWS answers what the service does with the bytes. You answer who may send them, where the copies land, and how long they stay. Walking a real feature through every copy it leaves behind shows how little of that second half survives choosing a Region and stopping there. On the professional track, the same properties get weighed against throughput and cost at a depth this level does not ask for.
Flash card
What Amazon Bedrock does with a prompt and the completion it returns, and what the customer still owns.
- AWS does not use your inputs or outputs to train Amazon Nova, Amazon Titan or any third-party model, and does not share them with the model provider. Retention is separate: nothing is kept by default, though a few models retain traffic inside AWS for 30 days for abuse detection.
- Prompts and completions travel over TLS 1.2 or better. Stored artefacts are encrypted at rest, and a custom model takes an AWS owned key unless you name a customer managed AWS Key Management Service [AWS KMS] key at job creation.
- An inference request is processed in the Region you call, the primary residency control. A cross-Region inference profile changes that: a geographic profile routes within a geography such as US or EU, a global profile to any commercial Region.
- Model invocation logging is disabled by default. Enable it and full prompts and completions go to an S3 bucket or CloudWatch Logs group in your account, bodies over 100 KB as separate S3 objects. Restrict who reads it, encrypt it, and set a retention period: CloudWatch Logs keeps data indefinitely.
- A fine-tuned model is a private copy only your account can invoke, stored and managed by AWS. The base model is unchanged and the training files are deleted when the job ends. It can reproduce training data in a completion, so filter confidential material out first.
Pick it when
Reach for these five lines when a scenario turns on whether a managed model is safe for regulated data. AWS runs the model, the hardware and the service encryption. You own who may invoke it, the key over any stored copy, the logging and the Region.
It's the wrong answer when
None of this makes a prompt safe to write. Bedrock not training on your text says nothing about personal details reaching the wrong reader, which is Amazon Bedrock Guardrails. Keeping the request off the public internet is AWS PrivateLink and a virtual private cloud [VPC] endpoint. AWS Artifact carries AWS compliance reports; CloudTrail records who called which model.