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
Q351 What is the MAIN limitation of relying only on transferring AI risk to a vendor?
- A. Risk is always fully removed once a contract assigns liability, so no further monitoring of the vendor is required
- B. Vendors never accept liability under any AI contract
- C. Accountability and reputational impact usually remain with the organization
- D. Regulators ignore contracts when assessing accountability
Answer: C
Transfer shifts some financial impact but accountability and reputational consequences typically stay with the organization. Risk is not removed, and vendors may accept liability in part.
Domain 2: AI Risk Management
Q352 Which statement BEST describes inappropriate risk?
- A. It is a risk that only affects vendors and never the enterprise, since the contract moves the consequences to the supplier
- B. It is always limited to cybersecurity events
- C. It is always acceptable if a vendor supplies the AI system
- D. It does not align with goals, produces no clear benefit or exceeds defined limits
Answer: D
Inappropriate risk does not align with the organization's goals, offers no clear benefit, or exceeds the defined risk tolerance. Appropriate risk is that necessary to pursue objectives.
Domain 2: AI Risk Management
Q353 A security architect builds an abstraction of an AI system, profiles likely attackers and catalogs potential threats. Which activity is MOST accurately being performed?
- A. Threat modeling
- B. Capacity planning
- C. Penetration testing
- D. Business impact analysis
Answer: A
Threat modeling creates an abstraction of the system, profiles attackers by goals, methods and skills, and catalogs potential threats. Penetration testing actively exercises controls, and a BIA measures the effect of loss. Capacity planning addresses resources.
Domain 2: AI Risk Management
Q354 When profiling a threat actor, which question is MOST useful for judging how likely an attack is to be attempted?
- A. Which cloud region stores archived logs and how long backups are retained
- B. Which font and layout the vendor uses in its public documentation set
- C. How many staff attend annual awareness training and how often policies are reissued
- D. What perceived value the asset has for the adversary and what effort is required
Answer: D
Threat modeling asks about perceived value, skills, time, resources and effort needed by the adversary, along with the danger of discovery. Documentation fonts, training attendance and log storage region do not describe attacker motivation.
Domain 2: AI Risk Management
Q355 A team applies STRIDE to an AI pipeline. What is the PRIMARY limitation to be aware of?
- A. It is too new to be understood and is not applied in industry practice
- B. It applies only to hardware devices rather than software and data flows
- C. It cannot identify data tampering or denial of service threats, which means it is unsuitable for any AI pipeline that needs threat analysis
- D. It does not by itself address challenges unique to AI such as data poisoning and adversarial machine learning
Answer: D
STRIDE is mature and easy to apply, helping identify tampering and denial of service, but it lacks the ability to address AI-specific issues such as adversarial ML and data poisoning. It is not limited to hardware, and it is an established method.
Domain 2: AI Risk Management
Q356 Why is it BEST for a team to combine several threat modeling methods for an AI system?
- A. Methods differ in scope, so combining them gives a more robust view of threats
- B. Combining methods removes the need to implement technical controls later
- C. Regulators require exactly three methods for any high-risk AI deployment
- D. Any single method is guaranteed to be complete, so combining adds only cost
Answer: A
Not all methods are comprehensive; some are abstract, others people-centric or focused on risk or privacy, so combining them gives a well-rounded view. No single method is fully complete, and threat modeling does not replace controls.
Domain 2: AI Risk Management
Q357 A control defends against both nation-state attackers and cybercriminals. How should the threat model BEST treat it?
- A. Remove it as redundant, since a control should address only one actor and duplicated coverage wastes budget
- B. Note it as an overlapping control that addresses more than one threat community
- C. Ignore it, since cybercriminals are outside the modeled scope
- D. Count it only once under the nation-state actor profile
Answer: B
Controls may address threats from more than one source, and overlapping controls should be noted in the threat model. Removing a control that serves several communities would reduce defense. Profiles are maintained per threat community.
Domain 2: AI Risk Management
Q358 Which set BEST describes the three major attack surfaces for AI threats identified by OWASP AI Exchange?
- A. Development-time threats, runtime security threats and threats through use
- B. Preventive, detective and corrective threats to the model pipeline, grouped by the control type that would address them
- C. Strategic, tactical and operational threats to the AI program
- D. Physical, logical and administrative threats across the AI estate
Answer: A
OWASP AI Exchange identifies development-time, runtime security and through-use threats as the main attack surfaces. The other groupings describe control classes or other taxonomies that are not specific to AI attack surfaces.
Domain 2: AI Risk Management
Q359 Which threat is MOST associated with the development-time attack surface?
- A. An end user typing a crafted prompt into a deployed assistant
- B. Network latency affecting the production inference endpoint
- C. Compromise of the engineering environment or data supply chain
- D. A mistyped query returning an inaccurate report to a manager
Answer: C
Development-time threats occur while the AI system is built, with the engineering environment and supply chain as attack surfaces. Crafted prompts are threats through use. Latency and mistyped queries are not threats to the development environment.
Domain 2: AI Risk Management
Q360 AI systems are also IT systems. Which threat MOST clearly falls under runtime application security threats?
- A. A poisoned training label introduced by a data supplier
- B. A vulnerability in the software wrapper or infrastructure hosting the model
- C. A biased training sample drawn from an unrepresentative population and used during an earlier training run
- D. A hallucinated answer produced after an ambiguous user prompt
Answer: B
Runtime application security threats stem from non-AI-specific weaknesses, such as vulnerabilities in the wrappers and infrastructure that host the solution. Poisoned labels and bias arise earlier, and hallucinations occur during use.