AIF-C01 security, compliance and governance
Security, Compliance, and Governance for AI Solutions is worth 14% of AIF-C01’s scored content — around 7 of the 50 scored questions. Like responsible AI, it is small, predictable and cheap to learn.
Remember the scope limit: AWS states that implementing security protocols and developing governance frameworks are out of scope for this candidate. You need to know the concepts and who is responsible for what. You do not need to configure anything.
The shared responsibility model applied to AI
The single most valuable hour of study in this domain, because it is a fixed concept that reliably yields marks.
AWS is responsible for security of the cloud: physical facilities, the hardware, the underlying infrastructure, and the managed service itself.
You are responsible for security in the cloud: your data, who can access it, how you configure the service, and what you choose to send to a model.
| Concern | Whose? |
|---|---|
| Data centre physical security | AWS |
| Patching the underlying infrastructure | AWS |
| The managed service’s own availability | AWS |
| Your training data and prompts | You |
| IAM permissions and who can call the service | You |
| Whether sensitive data is sent to a model | You |
| Encryption settings you control | You |
With managed AI services the line sits further toward AWS than with EC2 — but your data and your access control are always yours.
IAM
At foundational level:
- IAM controls who can do what to which AWS resources, including AI services.
- Least privilege — grant only what is needed.
- Roles over long-lived keys — an application should assume a role rather than carry a stored access key. This is the AWS equivalent of the managed-identity instinct, and it is usually the right answer.
Data protection
- Encryption at rest and in transit, and that AWS provides both.
- Where your data goes. A recurring concern with generative AI: what happens to prompts and inputs, and whether they could be used to train a model. Knowing this is a question worth asking is itself examinable.
- Data classification — knowing what is sensitive before deciding what may be sent to a model.
- Minimisation — not sending more than the task needs, the cheapest privacy control available.
Compliance
Concept level only:
- AWS holds compliance certifications and attestations; customers inherit some controls and remain responsible for their own use.
- Regulated industries and jurisdictions impose additional obligations, including on where data may be processed.
- Auditability — being able to show who did what, afterwards.
Governance
Also concept level, since developing frameworks is out of scope:
- An inventory of AI systems in use, including unapproved tools
- Ownership — a named accountable person
- Approval before deployment, proportionate to risk
- Monitoring after deployment
- Documentation of what a system does and its limitations
The recurring sequence, here as elsewhere: you cannot govern what you have not inventoried.
Sample questions
Question 1. Under the AWS shared responsibility model, who is responsible for deciding whether personal data may be included in prompts sent to a managed AI service?
- A. AWS, as the operator of the managed service
- B. The foundation model provider
- C. The customer
- D. The compliance auditor
Show answer
Answer: C
What data the customer chooses to send is a customer responsibility, since AWS secures the infrastructure and the service but does not decide what you submit. The model provider and a compliance auditor have no control over your submissions either.
Question 2. An application running on AWS needs to call a managed AI service. Security policy forbids storing long-lived credentials. What is the appropriate approach?
- A. Store an access key in an environment variable
- B. Have the application assume an IAM role with least-privilege permissions
- C. Store an access key in a configuration file with restricted permissions
- D. Rotate the access key every 24 hours
Show answer
Answer: B
Having the application assume an IAM role removes the stored credential entirely, which is what the policy requires. Storing a key in an environment variable or a configuration file still stores a credential, and rotating a key more frequently reduces exposure without eliminating it.
Question 3. An organisation discovers several teams have been using public generative AI tools with internal data, with no approval. Which governance step should come first?
- A. Inventory the AI tools in use and the data involved
- B. Block all public AI services at the network perimeter
- C. Deliver mandatory training to all staff
- D. Publish a policy prohibiting all external AI tools
Show answer
Answer: A
Establishing an inventory of what is in use and what data is involved must precede policy, blocking or training, because none of those can be designed sensibly without knowing the actual exposure. Blocking first also tends to drive usage underground.
What to practise
Draw the shared responsibility table from memory, then add three rows specific to AI: prompts, training data, and model output handling. If you can place those correctly, you can answer most of this domain — and the “roles not keys” instinct will carry you through the IAM questions on top.