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 2: AI Risk Management
Q371 A chatbot is tricked by a hidden instruction in user content into ignoring its guidelines. Which threat is MOST likely occurring?
- A. Model inversion
- B. Watermarking attack
- C. Data supply chain attack
- D. Prompt injection
Answer: D
Prompt injection manipulates the model through crafted inputs that override intended instructions. Inversion extracts data, a supply chain attack compromises components, and watermarking attacks target provenance marks.
Domain 2: AI Risk Management
Q372 An attacker uses AI to produce convincing voices and video to impersonate an executive. Which AI threat does this BEST represent?
- A. Model drift caused by changes in customer behavior over time
- B. Explainability failure caused by opaque model internals
- C. Bias amplification caused by skewed historical training data
- D. Deepfakes enabling identity and authentication abuse
Answer: D
Deepfakes generate synthetic media that can defeat identity checks and support social engineering. Drift, explainability failures and bias amplification are model quality or compliance issues.
Domain 2: AI Risk Management
Q373 A phishing campaign is generated at scale with personalized text from a language model. Which threat category does this BEST represent?
- A. Data poisoning
- B. Model extraction
- C. Infrastructure failure
- D. AI-generated social engineering
Answer: D
AI can create highly tailored social engineering campaigns. The other options describe attacks on models or availability. This is an AI-enabled threat rather than a threat to the model.
Domain 2: AI Risk Management
Q374 An AI loan model starts to favor one group more strongly than the training data showed. Which threat does this BEST describe?
- A. Model watermarking
- B. Bias amplification
- C. Denial of service
- D. Credential stuffing
Answer: B
Bias amplification occurs when a model exaggerates patterns in data, causing unfair outcomes. It falls under human factors and insider threats in the table. Credential stuffing and DoS are unrelated.
Domain 2: AI Risk Management
Q375 A model that was accurate at launch gradually gives worse predictions as real-world patterns change. Which threat is MOST relevant?
- A. Prompt injection
- B. Model drift
- C. Data exfiltration
- D. Model theft
Answer: B
Model drift is a change in model performance as data and conditions shift, under availability and reliability issues. Theft, injection and exfiltration are attacks, not gradual degradation.
Domain 2: AI Risk Management
Q376 A company cannot explain how its AI reached a decision that a regulator questions. Which threat category BEST applies?
- A. Cryptographic threats arising from weak key management in model storage
- B. Physical theft threats arising from loss of devices in transit
- C. Compliance and regulatory issues arising from explainability failures
- D. DoS threats arising from excessive adversarial query volume
Answer: C
Explainability failures create compliance and regulatory risk because decisions must be justified. Cryptographic and DoS threats are about integrity and availability. Physical theft is unrelated.
Domain 2: AI Risk Management
Q377 A third-party dataset used to fine-tune a model was tampered with by its supplier. Which threat category BEST applies?
- A. Insider misuse by the vendor's developers with privileged access
- B. Supply chain and third-party threats such as data supply chain attacks
- C. Availability threats arising from loss of the cloud hosting region
- D. Social engineering threats aimed at the help desk staff
Answer: B
Compromised datasets from suppliers are data supply chain attacks under supply chain and third-party threats. This is not an availability event or social engineering. The supplier tampering is a supply chain issue.
Domain 2: AI Risk Management
Q378 An employee uses an unapproved public AI service to summarize confidential documents. Which risk is MOST direct?
- A. Printer jams caused by additional document volume in the office
- B. Hardware wear on the servers that host the approved AI platform
- C. Office noise levels rising as more staff use voice assistants
- D. Unauthorized data disclosure through shadow AI
Answer: D
Using unapproved AI tools can move confidential data outside organizational control. Hardware wear and printer issues are unrelated. The risk is a data security and compliance one.
Domain 2: AI Risk Management
Q379 A security team wants to reduce the impact of model drift. Which mitigation is MOST appropriate?
- A. Replacing the perimeter firewall with a next-generation device
- B. Automated monitoring and model retraining pipelines
- C. Disabling logs to reduce the volume of monitoring noise
- D. Increasing network bandwidth between the data center and users
Answer: B
For drift, the manual lists automated monitoring and AI model retraining pipelines. Bandwidth and firewalls do not address performance changes, and disabling logs removes detection.
Domain 2: AI Risk Management
Q380 Which mitigation is MOST effective against deepfake misuse?
- A. More servers and wider load balancing for the inference service
- B. Watermarking, digital signatures, media forensics and public awareness
- C. Larger training datasets and additional compute for model training
- D. Longer retention of backups for the original media files
Answer: B
Deepfakes are mitigated by watermarking, digital signatures, media forensics and awareness. Bigger datasets or servers do not authenticate media. Backup retention is unrelated.