ISACA AAISM: Governing AI Security Risk

Governing AI security risk means deciding which AI-related exposures matter to the enterprise, how much risk the organization is willing to accept, which controls are proportionate, and who has authority to make those decisions. It is broader than securing a model. The risk can come from data, identity, providers, human use, autonomy, business process design, regulatory obligations, or concentration on a small number of external platforms.

The topic is the risk-management core of Enterprise AI Governance. In the current AAISM exam outline, AI risk management covers assessment, thresholds, treatment, threats, vulnerabilities, and vendor or supply-chain issues, while governance and technology domains connect those decisions to policies, controls, monitoring, incident response, and human oversight.

Translate AI uncertainty into scenarios

Statements such as ‘the model may hallucinate’ or ‘AI may expose data’ are too vague for governance. A scenario should describe the system, initiating condition, affected asset or stakeholder, business consequence, and control failure. For example: a customer-service assistant retrieves a restricted document and exposes account information to an unauthorized user.

Set appetite and tolerance for AI

Enterprise risk appetite should influence AI use. An organization may tolerate low-confidence drafting in an internal productivity tool while requiring much stricter evidence for automated access decisions or regulated advice. Tolerance should be expressed in thresholds that product and security teams can apply.

Prioritize by exposure, not novelty

Novel technology can attract disproportionate attention. A small internal chatbot may receive more review than a conventional automation with privileged access. Governance should compare risks based on impact, likelihood, detectability, concentration, and reversibility rather than whether the system contains a foundation model.

Govern vendor and supply-chain risk

AI services may depend on model providers, cloud platforms, data vendors, open-source libraries, vector databases, evaluation tools, and third-party agents. The risk assessment should identify which dependencies are critical, what data they receive, how they change, and whether the enterprise can monitor or replace them.