In AAISM, an adversarial attack is defined by maliciously crafted inputs designed to cause erroneous model behavior (misclassification, targeted misprediction) while the model architecture and parameters remain unchanged. Hardware compromise (A) and DoS (D) are system-level attacks, not adversarial ML per se.
Social engineering to obtain model information (C) is an information-gathering or security breach vector, distinct from input-space adversarial manipulation.
References:AI Security Management™ (AAISM) Body of Knowledge: Taxonomy of AI Threats- Adversarial Examples and Evasion; Distinction from System/Network Attacks.AAISM Study Guide:
Definitions and Threat Models in Adversarial ML; Attack-Defense Mapping for Model Predictions.

In AI governance frameworks, credit scoring is treated as a high-risk application. For such systems, the highest-priority safeguard is human oversight to ensure fairness, accountability, and prevention of bias in automated decisions.
The AI Security Management (AAISM) domain of AI Governance and Program Management emphasizes that high-impact AI systems require explicit governance structures and human accountability. Human-in-the- loop design ensures that final decisions remain the responsibility of human experts rather than being fully automated. This is particularly critical in financial contexts, where biased outputs can affect individuals' access to credit and create compliance risks.
Official ISACA AI governance guidance specifies:
High-risk AI systems must comply with strict requirements, including human oversight, transparency, and fairness.
The purpose of human oversight is to reduce risks to fundamental rights by ensuring humans can intervene or override an automated decision.
Bias controls are strengthened by requiring human review processes that can analyze outputs and prevent unfair discrimination.
Why other options are not the highest priority:
A). Regular updates improve accuracy but do not guarantee fairness or ethical decision-making. Model drift can introduce new bias if not governed properly.
B). Appeals mechanisms are important for accountability, but they operate after harm has occurred.
Governance frameworks emphasize prevention through human oversight in the decision loop.
D). Restricting criteria to "objective metrics" is insufficient, as even objective data can contain hidden proxies for protected attributes. Bias mitigation requires monitoring, testing, and human oversight, not only feature restriction.
AAISM Domain Alignment:
Domain 1 - AI Governance and Program Management: Ensures accountability, ethical oversight, and governance structures.
Domain 2 - AI Risk Management: Identifies and mitigates risks such as bias, discrimination, and lack of transparency.
Domain 3 - AI Technologies and Controls: Provides the technical enablers for implementing oversight mechanisms and bias detection tools.
References from AAISM and ISACA materials:
AAISM Exam Content Outline - Domain 1: AI Governance and Program Management (roles, responsibilities, oversight).
ISACA AI Governance Guidance (human oversight as mandatory in high-risk AI applications).
Bias and Fairness Controls in AI (human review and intervention as a primary safeguard).

Explanation:
Model distillation attacks involve using synthetic or carefully crafted data to query an AI model repeatedly and reconstruct its underlying parameters or behavior. This allows attackers to approximate the original model without direct access, effectively reverse engineering it.

Context awareness mechanisms help the AI system recognize malicious, irrelevant, or manipulative prompts that attempt to override instructions or alter intended behavior. This provides direct protection against prompt injection attacks by enabling the system to evaluate and constrain prompt context before responding.

AAISM risk management guidance clarifies that the organization's risk tolerance is the most important factor in determining how much monitoring is needed. Risk tolerance specifies the amount of risk the organization is willing to accept and defines the threshold for triggering monitoring or mitigation activities. Risk appetite is broader and strategic, while tolerance sets the operational limits. The number of users may influence scale, and compensating controls may affect resilience, but neither dictates monitoring intensity as directly as risk tolerance.