AIF-C01
Deep Dive
Five services cover most AI governance questions on the AIF-C01: who can access, how data is encrypted, who did what, what sensitive data exists, and what threats are detected against the infrastructure that supports the models.
Contents
| Service | Main function | Role in an AI context |
|---|---|---|
| AWS IAM | Access control and identity | Who can invoke models, create endpoints, or access training datasets. |
| AWS KMS | Encryption key management | Encrypts training data, model artifacts, and data in transit to the model. |
| AWS CloudTrail | API call auditing | Records who invoked which model, when, and with what result. |
| Amazon Macie | PII discovery and classification | Finds sensitive data in S3 before using it to train or give context to a model. |
| Amazon GuardDuty | ML-powered threat detection | Protects the infrastructure (accounts, network) that supports ML workloads. |
| AWS Config | Continuous configuration auditing | Evaluates whether the resources supporting your AI workloads (buckets, roles, endpoints) meet configuration rules. |
| Amazon Inspector | Automatic vulnerability scanning | Detects software and network vulnerabilities in the infrastructure that deploys your ML workloads. |
| AWS Audit Manager | Audit evidence automation | Continuously collects evidence to demonstrate regulatory compliance for your AI services. |
| AWS Artifact | Compliance reports portal | Access to AWS audit reports and agreements (needed to demonstrate compliance to third parties). |
| AWS Trusted Advisor | Automated recommendations | Includes security checks (for example, overly open permissions) relevant to AI environments. |
Controls access to AI services at the identity and permission level
In an AI context, IAM decides who can invoke a model in Bedrock, who can create or delete a SageMaker endpoint, and who can read training data stored in S3.
Typical use cases
Key principle
A role used by a chat app has excessive permissions over Bedrock. Find and fix the responsible statement — no guided steps.
+1 more...
Manages the keys that encrypt training data and models
KMS encrypts training datasets stored in S3, model artifacts (trained weights), and data in transit to and from an inference endpoint — just like it encrypts any other data on AWS.
Logs every invocation to Bedrock, SageMaker, and other AI services
CloudTrail answers "who invoked which model, when, and with what result?" — essential for compliance audits and for investigating an incident (for example, misuse of a model).
Enable a CloudTrail trail to audit who is invoking which Bedrock and SageMaker models, and from where.
+1 more...
Automatically classifies sensitive data in S3 with machine learning
Before using a dataset to train or connecting it as a context source for RAG, Macie can scan the S3 bucket and alert if it contains personal data (names, credit cards, IDs) that should be anonymized first.
For the exam: if the scenario mentions "verify sensitive data before training a model" or "PII in S3", the answer is Macie — not Comprehend (which analyzes text for sentiment/entities in general) nor GuardDuty (which detects threats, not content classification).
Threat detection for the accounts and network supporting your AI workloads
GuardDuty doesn't analyze models or prompts — it monitors anomalous behavior in the AWS account where your ML workloads run: access from malicious IPs, crypto-mining activity on instances that should be training models, or unusual API calls against your SageMaker resources. Its role is to protect the infrastructure around the ML pipeline, complementing IAM, KMS, and CloudTrail.
Five additional services the official guide explicitly names for governance and compliance of AI workloads — each covering a different part of the audit cycle.
AWS Config
Continuously evaluates and audits your AWS resources' configuration against compliance rules — for example, "do all buckets with training datasets have encryption enabled?". It also records the configuration change history.
Amazon Inspector
Automatically scans workloads (EC2 instances, container images, Lambda functions) for software vulnerabilities and unintended network exposure — relevant for the infrastructure that deploys inference endpoints.
AWS Audit Manager
Automates continuous evidence collection (configurations, logs, policies) to simplify compliance audits against frameworks like ISO 27001, PCI DSS, or internal frameworks.
AWS Artifact
Self-service portal to access AWS compliance reports (SOC, ISO, PCI DSS) and agreements like the BAA (HIPAA) — the source for demonstrating to an external auditor what AWS certifies.
AWS Trusted Advisor
Offers automated recommendations across several categories, including security — for example, IAM roles with excessive permissions or buckets with overly open access — checks that apply equally well to resources supporting AI workloads.
How to tell them apart on the exam
Config audits "resource configuration". Inspector looks for "vulnerabilities". Audit Manager collects "audit evidence". Artifact delivers "AWS's existing compliance reports and agreements". Trusted Advisor gives "general recommendations" (not just security). If the scenario mentions one of these four terms almost verbatim, that's the hint toward the right service.
Create a compliance assessment for a generative AI governance framework and gather evidence automatically.
+1 more...
An AWS framework that helps classify a GenAI use case into one of 5 "scopes" (categories) based on how much control you have over the model — from consuming a third-party application to training your own model from scratch. The scope determines which security and governance controls fall to you versus what the provider already covers.
Scope 1
Consuming a third-party SaaS application that uses GenAI internally — minimal control, almost all responsibility is the provider's
Scope 2
Building on top of a third-party application (e.g. a plugin or integration) — a bit more control and responsibility
Scope 3
Building your application using a pre-trained foundation model via API (e.g. Bedrock) — you control prompts, data, and guardrails
Scope 4
Fine-tuning an existing foundation model with your own data — more control, more responsibility over that data
Scope 5
Training your own model from scratch — maximum control and responsibility over the entire model lifecycle
The higher the scope (more control over the model), the greater the security and governance responsibility that falls on you — the same principle as the shared responsibility model, applied specifically to generative AI.
No guided steps: read a scenario and classify it into the correct scope of AWS's Generative AI Security Scoping Matrix.
+1 more...
Before taking an AI use case to production, check these five points.
Classify data sensitivity
Use Amazon Macie to discover whether your training or context data contains PII, and decide whether it needs to be anonymized.
Define who can access
Configure least-privilege IAM roles: who can invoke the model, who can deploy changes, who can read the data.
Configure Guardrails
Define harmful content filters, denied topics, and PII masking in the model's responses, appropriate to your use case.
Enable auditing
Turn on CloudTrail to log every invocation to AI services, and consider GuardDuty to monitor the underlying infrastructure.
Document the model with Model Cards
Record the purpose, training data, known limitations, and per-subgroup metrics before exposing the model to real users.
Define data residency
Decide in which region or country training and inference data is stored and processed — along with lifecycle, logging, monitoring, and retention policies, it's a central piece of data governance, especially if you operate under regulations requiring data not to leave a specific jurisdiction.
Did you get this topic?
Test what you just learned
Before connecting an S3 bucket as a source for a Bedrock RAG Knowledge Base, the security team wants to confirm the bucket does not contain personal ID numbers or credit card numbers. Which AWS service is best suited for this check?
Put what you learned to the test with original AIF-C01 exam questions.
Start AIF-C01 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.