Governance and compliance: AIF-C01 study notes and practice questions
What to know about Governance and compliance for the AIF-C01 exam, from the Security, Compliance, and Governance for AI Solutions domain (14% of the exam guide), then 2 free questions to test it. Part of the AIF-C01 study guide.
Compliance questions ask which AWS service gives a kind of evidence, or which framework helps govern AI.
AWS services for compliance and governance
Need
Service
AWS’s own compliance reports and agreements
AWS Artifact
Collect evidence that your AWS use meets a framework
AWS Audit Manager
Record resource configurations and check them against rules
AWS Config
Record who called which API
AWS CloudTrail
Find software vulnerabilities
Amazon Inspector
Check against best practices
AWS Trusted Advisor
Central security findings
AWS Security Hub
Governance practices
Policies for acceptable AI use, data handling and model approval.
Review cadence: regular reviews of models, data and outcomes.
Review strategies: technical, legal and ethics reviews before launch.
Transparency standards: document what models do and their limits (model cards, AI Service Cards).
Team training on responsible and secure use.
Frameworks such as the Generative AI Security Scoping Matrix, AWS’s model for matching security controls to how much of the AI stack you own.
Generative AI Security Scoping Matrix
It describes five scopes, from least to most ownership:
Consumer app (using a public generative AI service)
Enterprise app (a business application with generative AI features)
Pre-trained models (building on a model through an API such as Bedrock)
Fine-tuned models (customizing a model with your data)
Self-trained models (training your own from scratch)
The further along the scale, the more of the security, governance and compliance responsibility is yours.
Standards and regulations
ISO standards (such as ISO/IEC 42001 for AI management systems), SOC reports, and laws on data protection and AI. Algorithm accountability laws can require explanations and fairness testing.
Common traps
Artifact vs Audit Manager. Artifact gives AWS’s reports; Audit Manager gathers evidence about your own use.
Config vs CloudTrail. Config shows configurations; CloudTrail shows actions.
A healthcare company's governance team is setting up oversight for its AI workloads on AWS.
Match each need to the AWS service that meets it. Each option is used at most once, and one option is not needed.
Match each item to one answer. An answer can be used more than once, and some may not be used.
Items
Show auditors a record of which IAM users and roles called Amazon Bedrock APIs, and when
Download a SOC 2 report to share with a hospital customer's security reviewers
Keep a history of each S3 bucket's configuration and flag any bucket that breaks a rule, such as public access
Record each model's intended use, risk rating and evaluation results in one document for review
Answers
AA. AWS CloudTrail, which records API calls made by users, roles and services for auditing.
BB. AWS Artifact, which provides on-demand AWS compliance reports such as SOC reports.
CC. AWS Config, which records resource configuration history and evaluates it against rules.
DD. Amazon SageMaker AI Model Cards, which document a model's intended uses, risk rating and evaluation results.
EE. Amazon Bedrock Guardrails, which filter harmful content and sensitive information in model inputs and outputs.
Show the answer and explanation
Correct matches
Show auditors a record of which IAM users and roles called Amazon Bedrock APIs, and when: A. AWS CloudTrail, which records API calls made by users, roles and services for auditing.
Download a SOC 2 report to share with a hospital customer's security reviewers: B. AWS Artifact, which provides on-demand AWS compliance reports such as SOC reports.
Keep a history of each S3 bucket's configuration and flag any bucket that breaks a rule, such as public access: C. AWS Config, which records resource configuration history and evaluates it against rules.
Record each model's intended use, risk rating and evaluation results in one document for review: D. Amazon SageMaker AI Model Cards, which document a model's intended uses, risk rating and evaluation results.
About each answer
A. Matches the first need. CloudTrail records which identity made each API call and when, which is the audit trail auditors ask for.
B. Matches the second need. Artifact is where AWS compliance reports are downloaded for sharing with customers and reviewers.
C. Matches the third need. Config records how each resource's settings changed over time and flags resources that break a rule, such as public bucket access.
D. Matches the fourth need. Model cards hold a model's intended uses, risk rating and evaluation results in one record that can be exported.
E. Not needed here. Guardrails enforces safety at runtime, while the four needs are about audit records, compliance reports, configuration history and model documentation.
Explanation
Governance uses different AWS tools for different kinds of evidence. CloudTrail shows who did what, Config shows how resources were set up and changed, Artifact provides compliance reports, and Model Cards document each model. Guardrails enforces safety at runtime, so it is not one of the four needs. This question tests matching each kind of evidence to the right service.
A bank's governance team is mapping its generative AI projects to the AWS Generative AI Security Scoping Matrix to plan controls.
Put these scopes in order, from the LEAST ownership of the model to the MOST ownership.
Put these 5 steps in order.
AA. Building a model's weights from scratch on training data the bank owns or acquires.
BB. Using a public consumer chatbot that staff access through a web browser.
CC. Using a business software product with generative AI features built in, under an enterprise agreement.
DD. Calling a third-party foundation model through an API from the bank's own application.
EE. Further training a foundation model on the bank's own business data to create a specialized model.
Show the answer and explanation
Correct order: B, C, D, E, A
B. Using a public consumer chatbot that staff access through a web browser. First: a public consumer service gives the bank no view of the model or its training data, so its ownership is the lowest.
C. Using a business software product with generative AI features built in, under an enterprise agreement. Second: the bank uses an enterprise product with generative AI built in. It still does not control the model, but the vendor agreement gives it more say.
D. Calling a third-party foundation model through an API from the bank's own application. Third: the bank builds its own application on a third-party model. That adds control over the application, but the provider still owns and runs the model.
E. Further training a foundation model on the bank's own business data to create a specialized model. Fourth: the bank customizes an existing model with its own data, so it shares ownership of the model's behavior with the provider.
A. Building a model's weights from scratch on training data the bank owns or acquires. Fifth: the bank trains a model from scratch and owns every part of it, so it has the most ownership and the most responsibility for controls.
Explanation
The Generative AI Security Scoping Matrix describes five scopes, from using consumer apps to building and training your own model. Further along the scale, the organization owns more of the model, its data and its controls, and must secure more of them. This question tests ordering the scopes by ownership.
Want more on Governance and compliance? The 5 full-length practice sets have 11 more Governance and compliance questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →