Many companies hesitate to use generative AI (GenAI) because they worry about where their data goes, whether they can meet regulations, and what the model might say to customers. Running GenAI on AWS infrastructure answers those concerns with four benefits the exam names: security, compliance, shared responsibility, and safety. This lesson explains each one using Amazon Bedrock as the main example.
Security of Your Data
With Amazon Bedrock, your prompts and the model's responses are not used to improve the base models and are not shared with the model providers. Bedrock runs each provider's model in deployment accounts that the Bedrock service team owns and operates. The model providers have no access to those accounts, so they can't see customer prompts, completions or logs.
On top of that, the usual AWS security controls apply:
- Encryption of data in transit and at rest.
- AWS Identity and Access Management (IAM) to control exactly who and what can call which models.
- Amazon VPC and AWS PrivateLink: a VPC interface endpoint gives you a private connection to Bedrock, so traffic doesn't travel over the public internet.
- AWS CloudTrail to log API activity for auditing.
Compliance
AWS services are assessed by third-party auditors under many compliance programs. The Bedrock product page lists standards such as SOC, ISO, GDPR and FedRAMP High, and says it's HIPAA eligible. You can check which services are in scope for a given program on AWS's "services in scope" pages, and download the auditors' reports from AWS Artifact.
Being in scope helps, but it doesn't make your own application compliant automatically. How you configure and use the service is still up to you.
The Shared Responsibility Model
Security on AWS is split in two:
- Security of the cloud is AWS's job: the hardware, data centres, networking and the software that runs the managed service.
- Security in the cloud is the customer's job: your data, who has access to it, how services are configured, and how your application uses the model's output.
For a managed service like Bedrock, AWS runs much more of the stack than it does for a virtual server you manage yourself. But you still decide who gets IAM permissions, what data you send, and whether you turn on protections like guardrails.
Safety of Model Outputs
Amazon Bedrock Guardrails add configurable safeguards that work across different foundation models. A guardrail can filter harmful content (hate, insults, sexual content, violence, misconduct and prompt attacks), block topics you define, block specific words, and block or mask sensitive information such as personal details. It can also check whether answers are grounded in your source documents. One guardrail can protect several applications in the same consistent way.
On the Exam
| Concern in the scenario | AWS answer |
|---|---|
| "Will the model provider see or train on our data?" | Bedrock doesn't share prompts with providers or use them to train base models |
| "Traffic must not cross the public internet" | VPC interface endpoint with AWS PrivateLink |
| "We need the auditor's report" | AWS Artifact |
| "Block harmful replies and hide personal data" | Amazon Bedrock Guardrails |
| "Who handles what?" | Shared responsibility model |