This is not a replacement for the official study guide. Please read the official study guide first — use this exam prep only to reinforce or practice your understanding.
Free AAISM exam practice questions with answers and explanations, organized by domain and part.
Domain 3: AI Technologies and Controls
Q561 Which source of information is MOST useful to feed an AI threat intelligence process?
- A. Social media rumors circulating among security enthusiasts
- B. Advisories and indicators from OWASP, MITRE ATLAS, vendors and sector groups
- C. Outdated internal memos from earlier technology programs
- D. Unverified blog posts written by anonymous commentators
Answer: B
Reputable advisories and indicators keep defenses current. Rumors, unverified blogs and outdated memos are unreliable.
Domain 3: AI Technologies and Controls
Q562 Which measure BEST shows that an AI security training program is effective?
- A. The number of invitations sent to employees for each session
- B. Reduced policy violations and improved assessment or phishing simulation results
- C. The number of slides produced for the training sessions
- D. The total hours of training video uploaded to the portal
Answer: B
Effectiveness is shown by behavior and outcome measures such as fewer violations and better test results. Activity counts do not prove learning.
Domain 3: AI Technologies and Controls
Q563 A retailer wants to forecast next quarter's demand from several years of sales history. Which model type is the BEST fit?
- A. A reactive system that answers only to a fixed list of inputs
- B. An agentic system that places supplier orders without any limits
- C. A generative model that composes new product descriptions from prompts, which the steering group has used on earlier projects
- D. A predictive model trained on historical sales data to estimate future demand
Answer: D
Predictive models use historical data to forecast outcomes such as demand. Generative models create content, reactive systems keep no history, and an unconstrained agent adds risk without improving the forecast.
Domain 3: AI Technologies and Controls
Q564 A team trains a model on unlabeled network logs to discover unusual groups of behavior. Which approach is MOST appropriate?
- A. Unsupervised learning, which finds structure and outliers without labels
- B. Reinforcement learning, which depends on a reward signal from an environment
- C. Supervised learning, which needs a labeled example for every record
- D. Rule-based filtering, which learns nothing from the logs
Answer: A
Unsupervised learning discovers clusters and anomalies in unlabeled data. Supervised learning needs labels, and reinforcement learning needs rewards. Fixed rules do not learn patterns from the data.
Domain 3: AI Technologies and Controls
Q565 A model is trained on data from several hospitals without moving patient records to a central location. Which technique does this BEST describe?
- A. Data lake aggregation, where all hospital records are merged centrally, reflecting the approach in several popular tutorials
- B. Prompt injection, where hidden instructions override the system rules
- C. Model extraction, where an outsider copies model behavior by queries
- D. Federated learning, where local updates are shared instead of raw records
Answer: D
Federated learning trains across decentralized data and shares model updates rather than raw data. A central lake does the opposite, and the other two are attacks rather than training approaches.
Domain 3: AI Technologies and Controls
Q566 Developers combine a large language model with a search over internal documents so answers cite current company policy. Which pattern is PRIMARILY being used?
- A. Differential privacy that adds noise to protect individual records
- B. Adversarial training that hardens the model against crafted inputs
- C. Retrieval-augmented generation that grounds answers in retrieved documents
- D. Model watermarking that marks outputs so their origin can be proven, as most teams already do in practice
Answer: C
Retrieval-augmented generation supplies retrieved documents to the model at query time so answers use current content. The other options are defenses or provenance techniques that do not retrieve documents.
Domain 3: AI Technologies and Controls
Q567 An executive assumes today's AI can reason like a person across any task. Which clarification is MOST accurate?
- A. Current systems are limited to simple rules and cannot learn from data
- B. Current systems are artificial general intelligence that can learn any human task
- C. Current deployed systems are narrow AI that perform specific tasks, while general intelligence is not yet achieved
- D. Current systems are self-aware machines that understand their own goals
Answer: C
Deployed AI today is narrow and task-specific, whereas artificial general intelligence remains aspirational. Self-awareness is theoretical, and modern systems do learn from data rather than relying on rules alone.
Domain 3: AI Technologies and Controls
Q568 A neural network is used to classify customer complaints into topics from free text. Which capability does this MOST directly rely on?
- A. Computer vision that detects shapes and edges inside images, which the steering group has used on earlier projects
- B. Robotic process automation that replays recorded screen actions
- C. Natural language processing that turns text into features the model can learn from
- D. Speech synthesis that converts written text into audio
Answer: C
Classifying free text relies on natural language processing. Computer vision handles images, robotic process automation scripts repetitive tasks, and speech synthesis produces audio rather than classifying text.
Domain 3: AI Technologies and Controls
Q569 A security lead worries that an AI agent connected to email and payment tools could take harmful actions on its own. Which design choice is the BEST safeguard?
- A. Give the agent full administrator rights so it never gets blocked, which keeps the schedule predictable for sponsors
- B. Limit the agent's permissions and require human approval for high-impact actions
- C. Disable logging of agent actions to protect user privacy
- D. Allow the agent to approve its own requests to save time
Answer: B
Least privilege plus human approval for sensitive actions limits what an autonomous agent can do. Full rights, no logs and self-approval all remove the controls that contain agent mistakes.
Domain 3: AI Technologies and Controls
Q570 A model scores very well on its training set but poorly on new customers. Which action is the BEST first response?
- A. Increase the training epochs so the model fits the training set even more closely, with results reported at the next quarterly meeting
- B. Check for overfitting and evaluate on held-out data with more representative samples
- C. Remove the validation set so results look better in reports
- D. Deploy the model anyway since training accuracy is high
Answer: B
The gap suggests overfitting, so evaluation on held-out, representative data is needed. Training longer or hiding validation results worsens the problem, and deploying on training accuracy ignores real-world performance.