AI safety, security, and governance
AWS Certified Generative AI Developer - Professional exam topic, about 20% of the exam. Updated .
Keep the model useful and on a leash
A model will cheerfully do what the prompt says, including what a stranger smuggled into the prompt. This domain is about not letting that happen, and about proving later that it did not.
Safety controls, in layers
Amazon Bedrock Guardrails is the centrepiece. It applies content filters (hate, insults, sexual content, violence, misconduct, and a prompt attack filter for jailbreaks and injection), denied topics, word filters, sensitive information filters for PII, and a contextual grounding check that scores answers against the source passages. The ApplyGuardrail API lets you run the same checks on text that never touched Bedrock, such as another provider's output.
No single control is enough, so build defence in depth. Pre-process with Amazon Comprehend or Lambda checks, apply the guardrail on input and output, validate structure afterwards (JSON Schema, allow-lists), and log. Reduce hallucinations by grounding answers in a Knowledge Base, citing sources, and refusing when retrieval finds nothing. For text-to-SQL, constrain the model to read-only, parameterised queries so results stay deterministic and safe.
Data security and privacy
Keep traffic private with VPC interface endpoints (PrivateLink) for bedrock-runtime and bedrock-agent-runtime, control access with least-privilege IAM, and encrypt with KMS. Bedrock does not use your prompts or outputs to train base models. Find PII with Amazon Comprehend and Amazon Macie, mask it before it reaches the model, and set S3 lifecycle rules so data does not live forever.
Governance and responsible AI
Governance means traceability. CloudTrail records who called what, model invocation logging captures prompt and response content when you enable it (it is off by default), and Glue Data Catalog and metadata tags track where data came from. Model cards document limits. Keep decision logs in CloudWatch Logs and alarm on misuse patterns.
Responsible AI adds transparency and fairness: show sources, expose agent traces where appropriate, and test for bias with evaluations and A/B tests. Pass the audit by being boring and well documented.