ISACA AAISM: AI Model Risk for Security Leaders
An AI model can become a security risk even when it never suffers a conventional software exploit. Sensitive training data may be memorized, a retrieval system may disclose information across tenants, a model or prompt can be manipulated, and an automated action can reach a system it should not control. Security leaders therefore need an AI model risk record that connects technical behavior to business impact. ISACA’s AAISM exam outline covers AI governance, AI risk management and AI technologies and controls; the following method translates those themes into practical review.
On this page
Inventory the whole AI application
Describe the model, deployment environment, upstream data, retrieval sources, prompts, evaluation sets, tools, actions, identities and service providers. Include the business process the system supports and the intended user population. A model identifier alone is insufficient, especially when external APIs or agents can access organizational systems. Identify data residency, retention, contractual limits and the point where human approval is supposed to occur. Mark which components are controlled internally and which are supplier responsibilities.
Write risks as testable failure scenarios
Replace generic statements such as “hallucinations are dangerous” with measurable cases. Example: a support assistant answers a customer question using another customer’s confidential ticket because retrieval authorization is not applied at document level. List the initiating condition, affected data or asset, consequence, likely detection method and owner. Distinguish a privacy breach, a wrong business recommendation, a service outage and an unauthorized tool action; they call for different treatments.
Set limits before selecting mitigations
Define what the model is permitted to do, how much automation can occur without review and what level of residual risk the business will accept. For a public-facing assistant, a low-stakes summarization error may be tolerable while sending a payment or changing account permissions without approval is not. Require independent enforcement in identity, application code and downstream APIs; a system prompt is not a security boundary. Track supplier restrictions, geographic controls and access to training or telemetry data.
Measure behavior under realistic adversarial and ordinary inputs
Build evaluation sets with representative user tasks, edge cases, ambiguous questions, malicious instructions embedded in retrieved content, and requests from principals with different access levels. Check factual grounding, refusal behavior, leakage, tool-use boundaries, latency and business-specific quality thresholds. Record model version, prompt version, retrieval corpus and tool permissions so results can be reproduced. A single demonstration pass cannot establish production safety; monitor drift and retest after material changes.
Govern exceptions and incident response
Assign a risk owner and time-bound review date to every exception. Specify kill switches, fallback workflows, evidence retention and escalation paths for suspected leakage or harmful actions. If a third-party model changes silently, the response may require additional monitoring or temporary reduction in tool privileges. Keep the AI asset inventory linked to change management and supplier contracts. Security and business stakeholders should sign off on residual risk with evidence, not on a vague promise of “responsible AI.”
What an AAISM candidate should be able to explain
Describe how model risks differ from ordinary infrastructure vulnerabilities, how governance and risk assessment connect to controls, and how to prove a safeguard works. Practice presenting one risk scenario to legal, product and operations stakeholders, each of whom needs a different decision. ISACA’s current eligibility and exam rules belong on its official credential page; this article is about applying risk-management judgment rather than providing unverified pass-rate or employment promises.