AAISM materials identify human-in-the-loop governance as the most effective safeguard against risks such as hallucinations in AI systems used in high-stakes domains like healthcare. By ensuring that human experts validate outputs before they influence patient treatment decisions, organizations preserve accountability, safety, and accuracy. Penetration testing is a cybersecurity measure, not relevant to hallucination risk. AI impact analysis helps evaluate systemic effects but does not directly prevent faulty outputs. Data validation improves input quality but cannot fully prevent generative hallucinations. The key safeguard is human-in-the- loop oversight.
References:
AAISM Study Guide - AI Governance and Program Management (Human Oversight in High-Risk AI) ISACA AI Security Management - Mitigating Hallucinations in Generative AI

AAISM categorizes the manipulation of an LLM at inference time, where crafted inputs cause outputs to serve attacker objectives, as an evasion attack. Evasion attacks exploit weaknesses in the model's decision- making boundaries by altering queries to produce compromised or misleading outputs. Privilege escalation refers to unauthorized access rights, data poisoning targets the training phase, and model inversion reconstructs training data. In this case, manipulation of outputs to align with an attacker's goals reflects an evasion attack.
References:
AAISM Exam Content Outline - AI Risk Management (Adversarial Attack Types) AI Security Management Study Guide - Evasion and Manipulation Risks

Themost effective SOC applicationof AI is indetecting subtle, hard-to-find attack patternsthat reduce false negatives.
AAISM technical control guidance notes that AI in SOCs is best applied to:
* Enhance detection accuracy and sensitivity to anomalies.
* Assist analysts in identifying hidden patterns that traditional rule-based systems miss.
* Augment-not replace-human decision-making for high-confidence outcomes.
Options B and C incorrectly shift responsibility entirely to AI, which contradicts governance principles requiringhuman oversight. Option D is useful for efficiency, but theprimary effectivenesscomes from improving detection quality.
Therefore, the most effective use is toreduce false negatives and detect subtle attacks.

AAISM defines adversarial attacks as manipulations of input data (text, image, audio, numeric values) designed to cause the model to produce incorrect or harmful predictions.
Hardware attacks (A) are infrastructure threats. Social engineering (C) targets people, not models. DoS attacks (D) affect availability, not model decision pathways.
References: AAISM Study Guide - Adversarial Threats; Input Manipulation.

An AI impact assessment prioritizes alignment with stated business objectives and intended purpose of the AI use case. Establishing whether the AI system measurably supports defined objectives (value, outcomes, success criteria) is the primary lens before assessing secondary concerns (workforce effects, awareness initiatives, reputational considerations). Governance requires confirming purpose limitation, success metrics, and risk-benefit trade-offs tied to organizational objectives; only then are downstream impacts evaluated.
References: AI Security Management (AAISM) Body of Knowledge - AI Governance & Impact Assessments; Purpose and Objective Definition; Risk-Benefit Evaluation. AAISM Study Guide - Business Alignment, Intended Use, and Success Criteria in AI Impact Assessments.