AAISM Practice Questions: AI Risk Management Domain

Correct Answer: A

Explanation: The best initial step in assessing the risk of adversarial attacks is conducting a comprehensive threat modeling exercise (Option A). This process helps identify potential threats, vulnerabilities, and attack vectors specific to the AI model. While penetration testing (Option B) and reviewing the architecture (Option C) are valuable, they should follow the threat modeling to ensure a targeted approach. Anomaly detection (Option D) is a control mechanism rather than an assessment step.

Correct Answer: B

Explanation: The best strategy for managing the risk of prompt injection attacks is implementing strict input validation and sanitization. This helps ensure that inputs are properly checked and sanitized before being processed by the AI system, reducing the risk of malicious inputs causing unexpected behavior. While updates, isolation, and audits are beneficial, they do not specifically address the risk posed by prompt injection attacks.

Correct Answer: B

Explanation: The potential impact of AI incidents on organizational operations is the most critical factor, as it directly affects the organization's ability to function and meet its objectives. While technological advancements, expert availability, and compliance are important, the primary concern is mitigating operational impact.

Correct Answer: B

Explanation: Implementing a robust audit trail for AI model decisions ensures transparency and explainability by allowing stakeholders to trace and understand how decisions are made. While updating models (A), using open-source frameworks (C), and providing training (D) are beneficial practices, they do not directly address the transparency and explainability of decision-making processes.

Correct Answer: D

Explanation: The first step should be to engage in dialogue with the vendor to understand their compliance posture. This approach allows the security manager to gather more information and potentially resolve the issue without escalating it unnecessarily. Immediate termination or notifying authorities should be considered only after confirming non-compliance and if the vendor is unwilling to address the issues.

Correct Answer: A

Explanation: Data poisoning involves injecting malicious data into the training dataset to corrupt the model. Implementing robust data validation and cleaning processes is the most effective control to identify and remove such malicious data before it can influence the model. While encryption and access controls are important, they do not directly address the integrity of the training data.

Correct Answer: A

Explanation: The sensitivity of the data used for training the model is the most critical factor because model inversion attacks aim to reconstruct sensitive data from the model outputs. The complexity of the model (B), data storage location (C), and update frequency (D) are less directly related to the risk of inversion attacks.

Correct Answer: B

Explanation: The sensitivity and value of the data being processed is the most critical factor when assessing the risk of adversarial machine learning attacks. High-value data increases the incentive for attackers to target AI systems. While model complexity (A), geographical location (C), and personnel availability (D) are relevant, they are secondary to the data's sensitivity and value.

Correct Answer: A

Explanation: Implementing differential privacy techniques is the most effective control for preventing model inversion attacks, as it adds noise to the data to protect individual privacy while maintaining the utility of the model. Regular updates, threat modeling, and access controls are important security practices but do not specifically address the risk of model inversion.

Correct Answer: B

Explanation: Conducting regular audits of AI vendors' security practices is the most effective way to manage AI supply-chain risks, as it provides ongoing assurance that vendors are maintaining adequate security measures. While contractual agreements, adherence to standards, and diversification are important, they do not provide the same level of continuous oversight.