AIF-C01 - Responsible and secure AI (12% of the exam) - Section 5.3
Explain methods to secure AI systems and meet governance requirements.
Apply least-privilege access, data protection, and prompt-injection defences, and recognise the governance and compliance obligations that apply to AI systems.
Prompt injectionData protectionAI governance
Practice question for this objective
lock_openFree sampleResponsible and secure AImedium
A healthcare organisation is deploying a generative AI assistant that may receive sensitive patient queries. The security team requires that data submitted to the model is NEVER used by the model provider to retrain their models. Which Amazon Bedrock feature addresses this requirement?
AActivating Amazon Bedrock Guardrails to filter harmful content from both inputs and outputs before they reach the underlying model.
BConfiguring a VPC endpoint (PrivateLink) for Amazon Bedrock so that all traffic stays within the AWS network and never traverses the public internet.
CUsing Amazon Bedrock's data privacy commitment, which ensures that customer prompts and responses are not used to train or improve AWS or third-party foundation models.check_circle Correct
DEnabling server-side encryption with a customer-managed KMS key on all Bedrock invocations to retain cryptographic control over model input data.
Understand that Amazon Bedrock's data privacy commitment ensures customer invocation data is not used to train foundation models, satisfying healthcare data governance requirements. Amazon Bedrock is designed so that customer inputs and model outputs are not shared with model providers and are not used to train or improve any model. This is a service-level commitment documented in AWS's data privacy terms, distinguishing Bedrock from some other AI APIs. Network controls and encryption address different threats and do not substitute for this data-use policy.
Why A is wrong: Guardrails filter content for safety and policy compliance but do not govern whether the model provider uses invocation data for retraining. Content filtering and data usage policies are separate concerns.
Why B is wrong: A VPC endpoint keeps traffic off the public internet and reduces network-level exposure, but the data still reaches the Bedrock service. Network routing does not affect the contractual or technical mechanisms governing whether prompt data is used for model training.
Why C is correct: AWS explicitly commits that data submitted to Amazon Bedrock - including prompts and completions - is not used to train the underlying foundation models. This is a contractual and architectural guarantee built into the service, directly addressing the organisation's requirement without any additional configuration.
Why D is wrong: Customer-managed KMS keys give the customer control over who can decrypt their data, which is a confidentiality and access-control measure. It does not prevent the model provider from processing the plaintext during inference or address the data-use policy for training.
Answers that look right on this material and are not. Each one is a distractor from a different question in the AIF-C01 bank for this domain.
Apply a single global standard set to whichever jurisdiction is most permissive, to keep the rollout simple and consistent
errorWhy it is wrong: Standardising on the most permissive jurisdiction is operationally simple and therefore tempting, but it leaves the high-risk jurisdiction non-compliant, which exposes the company to enforcement and penalties.
Raise the model's sampling temperature so its replies to customers vary more from one request to the next.
errorWhy it is wrong: Temperature only changes output randomness and has no bearing on whether a malicious instruction triggers the refund tool.
Enable AWS CloudTrail management event logging for Bedrock API calls, then archive CloudTrail logs to an S3 bucket with S3 Versioning enabled.
errorWhy it is wrong: CloudTrail logs management-plane API calls (e.g. ListFoundationModels) but does not capture the content of model invocations - the prompts and responses. S3 Versioning retains previous versions but does not make objects immutable; a user with delete permissions can still remove them.