Exam Prep

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

Q51 An AI research team is developing a natural language processing model that relies on several open-source libraries. Which of the following is the team’s BEST course of action to ensure the integrity of the software packages used?

Answer: B

AAISM’s technical control guidance emphasizes that when using open-source libraries, the best safeguard for integrity is to scan the packages for malware before installation. This ensures that compromised or malicious code does not enter the AI system environment. Maintaining lists aids consistency but not security. Always using the latest versions may introduce unverified vulnerabilities. Retraining models addresses functionality but not software integrity. Therefore, the strongest protective measure is pre-installation malware scanning of open-source packages.

Domain 3: AI Technologies and Controls

Q52 AI developers often find it difficult to explain the processes inside deep learning systems PRIMARILY because:

Answer: D

Deep learning models learn high-dimensional, non-linear representations through layered parameterization that resists simple causal narratives. The internal mechanisms (e.g., distributed feature representations and complex statistical transformations) are difficult to map to human-interpretable rules, making explanation challenging. This is the primary reason for explainability difficulty in deep learning. Option A addresses data origin/volatility, not explainability. Option B mischaracterizes model behavior as mutable “knowledge” without logs. Option C notes probabilistic foundations but that alone does not make systems inexplicable.

Domain 2: AI Risk Management

Q53 During the creation of a new large language model (LLM), an organization procured training data from multiple sources. Which of the following is MOST likely to address the CISO's security and privacy concerns?

Answer: B

AAISM guidance highlights data minimization as a critical practice for addressing both security and privacy concerns. By ensuring that only the minimum necessary data is collected and retained, the organization reduces the risk of sensitive information being exposed or misused during training. Data augmentation expands data but does not mitigate privacy risk. Classification organizes data but does not limit exposure. Data discovery helps locate sources but does not directly reduce risks. The control that directly aligns with privacy-by-design principles is data minimization.

Domain 3: AI Technologies and Controls

Q54 An attacker crafts inputs to a large language model (LLM) to exploit output integrity controls. Which of the following types of attacks does this BEST represent?

Answer: A

According to the AAISM framework, prompt injection is the act of deliberately crafting malicious or manipulative inputs to override, bypass, or exploit the model’s intended controls. In this case, the attacker is targeting the integrity of the model’s outputs by exploiting weaknesses in how it interprets and processes prompts. Jailbreaking is a subtype of prompt injection specifically designed to override safety restrictions, while evasion attacks target classification boundaries in other ML contexts, and remote code execution refers to system-level exploitation outside of the AI inference context. The most accurate classification of this attack is prompt injection.

Domain 2: AI Risk Management

Q55 Which of the following AI data life cycle phases presents the GREATEST inherent risk?

Answer: D

The data Preparation phase—covering sourcing, collection, labeling, cleansing, and provenance—presents the greatest inherent risk because it is where privacy, consent, representativeness, bias, quality, lineage, and legality must be established. Decisions and defects here propagate into training and downstream use, amplifying ethical, regulatory, and accuracy risks. While Training can introduce additional risks (e.g., poisoning, leakage), these are frequently mitigated by controls that depend on having trustworthy prepared data. Monitoring and Maintenance are later-life-cycle phases oriented toward detection and correction; they are critical but inherently rely on the foundation set during Preparation.

Domain 1: AI Governance and Program Management

Q56 During the deployment of a generative AI platform, a risk assessment highlighted threats such as data leakage and prompt manipulation. Which of the following is the BEST way to ensure appropriate control selection?

Answer: B

AAISM requires that control selection be threat-led and context-specific, aligning AI threats to the organization’s existing enterprise control catalogs (security, privacy, resilience) and augmenting them with AI- specific safeguards where coverage is insufficient. This ensures consistency with the risk appetite, removes duplication, and closes AI-unique gaps (e.g., prompt injection, data leakage from context windows, model misuse). Generic reliance on vendors or uncustomized external frameworks does not ensure fit-for-purpose coverage, and deferring control selection to post-deployment contradicts proactive risk treatment.

Domain 2: AI Risk Management

Q57 Which of the following will BEST reduce data bias in machine learning (ML) algorithms?

Answer: C

AAISM guidance clearly states that the most effective way to mitigate data bias is through diverse training data that fairly represents all relevant populations, scenarios, and contexts. Simplified models may reduce complexity but do not remove bias. Unstructured data sets may introduce new errors without addressing fairness. Securing training data protects confidentiality and integrity but does not resolve representational imbalance. Therefore, the best practice for reducing bias in ML is diversification of training datasets.

Domain 1: AI Governance and Program Management

Q58 A large financial services organization is integrating a third-party AI solution into its critical fraud detection system. Which of the following is the BEST way for the organization to reduce risk associated with AI vendor and supply chain dependencies?

Answer: C

AAISM emphasizes supplier assurance through contractual obligations as the foundational control for AI supply chain risk. Contracts should require verifiable evidence of secure development practices (e.g., secure SDLC, model and data provenance documentation, SBOM/MBOM where applicable, vulnerability disclosure, patch SLAs, audit rights, incident notification, and regulatory compliance assertions). This creates enforceable, continuous assurance beyond point-in-time tests. • A is necessary but reactive and limited to your environment. • B addresses performance, not supply chain security. • D is a good isolation/validation practice but does not create vendor accountability across the lifecycle.

Domain 3: AI Technologies and Controls

Q59 A data scientist creating categories and training the algorithm on large data sets BEST represents which type of AI model learning technique?

Answer: D

AAISM classifies learning paradigms by the presence of labeled targets. Creating categories (labels) and training on them is supervised learning, where input features are mapped to known outputs and optimization minimizes prediction error against ground truth. Unsupervised (B) discovers structure without labels; reinforcement (A) optimizes behavior via rewards; “machine learning” (C) is the broad field, not the specific technique.

Domain 2: AI Risk Management

Q60 When deriving statistical information from AI systems, which source of risk is MOST important to address?

Answer: D

AAISM emphasizes systemic or structural bias as a high-impact risk because biased data leads directly to discriminatory insights or decisions when used for analytics or reporting. This risk affects fairness, compliance, and organizational reputation. Hallucinations (A) relate more to generative AI. Incomplete outputs (B) affect accuracy but not structural fairness. Lack of normalization (C) affects performance but is not the dominant risk.

‹ Prev
Next ›
Page 6 of 70 · 700 questions