Customize this training

Get a certificate — $16.67$5

ISACA AAISM Exam Prep: Advanced in AI Security Management Course and Mock Exam

ISACA AAISM Exam Prep: Advanced in AI Security Management Course and Mock Exam

Revise the three AAISM domains: AI governance and program management, AI risk management, and AI technologies and controls. Then practise ISACA-style BEST and FIRST questions on a mock exam.

What you will learn

  • Explain the AAISM exam format and apply stakeholder needs, industry frameworks and regulations to AI governance.
  • Build AI strategy, policies, asset and data inventories, a security program and AI-ready incident response.
  • Assess AI risk, set risk thresholds and choose a suitable treatment for AI use cases.
  • Manage AI threats, vulnerabilities, vendors and the AI supply chain.
  • Apply secure architecture and design across the AI life cycle, from model selection to validation.
  • Select data management, privacy, ethics, trust and safety controls for AI systems.
  • Operate security controls and monitoring for AI systems in production.

Modules

  1. The AAISM exam, stakeholders, frameworks and regulation
  2. AI strategy, policies, assets, program and incident response
  3. AI risk assessment, thresholds and treatment
  4. AI threats, vulnerabilities, vendors and supply chain
  5. AI security architecture and the AI life cycle
  6. Data management, privacy, ethics, trust and safety controls
  7. Security controls and monitoring for AI in production

Course facts

Level
Exam preparation
Estimated study time
about 2 hours
Final assessment
42 questions · Pass mark 70%
Available in
English
Price
Free. Optional verifiable certificate: $5
Last updated
Published by
AMAADOR ACADEMY

Who this course is for

CISM and CISSP holders, security managers, CISOs, GRC leads and security architects who are preparing for ISACA AAISM or must run security for AI systems in their organisation.

Prerequisites

ISACA requires an active CISM or CISSP to sit AAISM. You should already know information security management (governance, risk, incident response, controls). No coding is needed.

Course content

1. The AAISM exam, stakeholders, frameworks and regulation

AAISM (Advanced in AI Security Management) is an ISACA credential for security managers who already hold CISM or CISSP. It does not re-test basic security. It tests whether you can extend an existing security program to cover AI. The facts below come from ISACA’s AAISM page and the AAISM Exam Content Outline, checked in November 2026.

ItemWhat ISACA states
EligibilityActive CISM or CISSP; eligibility lasts six months from registration
Questions90
Task statements22, across 3 domains
LanguagesEnglish, Chinese, Spanish, Japanese
Exam feeUS$459 members, US$599 non-members, plus a US$50 application fee
Outline changeA new outline takes effect on 18 May 2027
DomainWeightModules here
1. AI Governance and Program Management31%1-2
2. AI Risk Management31%3-4
3. AI Technologies and Controls38%5-7

This course teaches the outline in force until 17 May 2027. If you sit the exam on or after 18 May 2027, check the new outline first.

The ISACA mindset

  1. Read the qualifier — BEST, FIRST, MOST important, PRIMARY. Several options can be true; only one fits the qualifier.
  2. Think like the manager — you advise, align to business goals and make sure risk is owned. You do not tune models yourself.
  3. Governance before technology — when there is no policy, owner or risk assessment yet, the answer is usually to create that first.
  4. Risk-based, not fear-based — banning AI outright is rarely BEST; enabling it with controls inside the risk appetite usually is.

2. AI strategy, policies, assets, program and incident response

An AI security strategy starts from the business strategy. If the company plans to use AI for customer service and fraud detection, security must enable those uses safely. A strategy that only says no will be bypassed by shadow AI.

The policy stack

LayerExample contentOwner
AI policyPrinciples, risk appetite, roles, who may approve AI useSenior management, approved by the board or committee
Acceptable use standardWhich AI tools staff may use, what data may never be enteredCISO with legal and HR
ProceduresAI use case intake, model release checklist, red team processProcess owners
GuidelinesPrompt hygiene tips, labelling AI-generated contentSecurity awareness team

Governance structures matter more than documents. Many organisations create an AI governance committee with business, legal, privacy, risk, security and data science. It approves high-risk use cases and keeps the inventory current.

Do

  • Align the AI policy with existing security, privacy and data policies
  • Name an accountable owner for each AI use case
  • Provide an approved tool so staff have a safe option
  • Review the policy when regulation or technology changes

3. AI risk assessment, thresholds and treatment

AI risk assessment uses the same logic as any risk assessment: identify, analyse, evaluate, treat, monitor. What changes is the list of things that can go wrong and who can be harmed.

Risk categoryExample
SecurityPrompt injection, model theft, poisoning, data leakage
ReliabilityHallucination, drift, poor performance on new populations
FairnessSystematically worse outcomes for a protected group
PrivacyModel memorises and reveals personal data
Legal and complianceUse falls into a prohibited or high-risk category
Third partyProvider changes the model, terms or data use

NIST AI RMF calls the context step Map: understand the purpose, the users, the people affected, the data and the setting before you Measure. Many failures come from using a sound model in the wrong context.

  1. Describe the use case — purpose, decision supported, degree of autonomy
  2. Classify — risk tier using the regulatory and internal criteria
  3. Identify risks — use threat models such as OWASP and MITRE ATLAS plus impact on people
  4. Analyse — likelihood and impact, including impact on individuals and society
  5. Evaluate — compare with risk appetite and thresholds
  6. Record — risk register entry with an owner

4. AI threats, vulnerabilities, vendors and supply chain

AI systems keep every classic vulnerability and add new ones. The exam expects you to recognise the main attack types and to manage them through the normal vulnerability process.

ThreatWhat happensKey controls
Prompt injection (direct or indirect)Instructions in user input or in a document, web page or email override the system’s intended behaviourTreat all input as untrusted, least privilege for tools, output checks, human approval for actions
Data poisoningAttacker alters training or fine-tuning data to plant a backdoor or biasData provenance, integrity checks, validation on clean test sets
Evasion (adversarial examples)Crafted inputs make the model misclassifyAdversarial testing, input preprocessing, ensemble checks
Model extraction or theftCopying weights, or rebuilding a model by querying itAccess control, rate limits, monitoring of query patterns
Membership inference and inversionLearning whether a person was in training data, or recovering itData minimisation, differential privacy, output limits
Excessive agencyAn agent with broad tool access is tricked into harmful actionsNarrow permissions, confirmation steps, logging

OWASP’s Top 10 for LLM Applications (2025 edition) lists prompt injection first, followed by sensitive information disclosure and supply chain. MITRE ATLAS maps real adversary techniques against AI and is useful for red team plans.

5. AI security architecture and the AI life cycle

Security by design applies to AI. The difference is that the model itself is an untrusted component: it can be manipulated through its inputs and it produces outputs that may be wrong or malicious.

A typical generative AI application

  1. User and identity layer — authentication, role-based access, session limits
  2. Input guardrails — filtering, classification of sensitive data, size limits
  3. Orchestration — builds the prompt, calls tools and retrieval
  4. Retrieval (RAG) — vector database of company documents; enforce the user’s own permissions on what is retrieved
  5. Model — hosted, API or self-hosted; isolated, with no direct access to secrets
  6. Output guardrails — check for sensitive data, unsafe content, and treat output as untrusted before passing it to code, browsers or other systems
  7. Logging and monitoring — prompts, outputs, tool calls, with privacy rules

Do

  • Apply least privilege to every tool and data source the model can reach
  • Keep secrets and keys out of prompts and system prompts
  • Enforce document-level access control in retrieval
  • Segment training, development and production environments

Don't

  • Rely on the system prompt as a security boundary
  • Execute model output as code or SQL without validation
  • Give a single service account access to all documents for RAG

6. Data management, privacy, ethics, trust and safety controls

Data is the largest attack surface in AI. Training data, fine-tuning data, retrieval documents, prompts, outputs and embeddings all need classification and protection.

ControlWhat it does for AI
Classification and labellingDecides which data may be used for training, retrieval or external tools
Lineage and provenanceRecords where data came from and how it changed; supports poisoning investigations
Quality checksFinds errors, duplicates, gaps and unrepresentative samples before training
Integrity protectionHashes, versioning and restricted write access to training sets
MinimisationUses only the data needed; removes direct identifiers where possible
Retention and deletionApplies retention to prompts, logs, embeddings and model copies
Encryption and access controlProtects data at rest, in transit and in vector stores

Embeddings are not anonymised data. Research has shown that text can sometimes be reconstructed from embeddings, so vector databases deserve the same classification as their source documents.

Do

  • Keep a data sheet for every training dataset
  • Separate test data from training data and lock it
  • Apply DLP to prompts sent to external AI services

7. Security controls and monitoring for AI in production

Once a model is live, the classic control families still apply. The exam tests whether you can apply each to AI assets.

Control familyAI application
Identity and accessSeparate roles for model developers, operators and users; MFA for model registries; scoped API keys
Change managementEvery model update, prompt change or new tool is a change with testing and approval
Configuration and hardeningLocked system prompts, disabled unused tools, safe default parameters
Resource protectionRate limits and quotas against abuse and runaway costs (unbounded consumption)
Secrets managementKeys in a vault, never in prompts, notebooks or code repositories
Backup and recoveryVersioned models, data and configurations; tested rollback
AwarenessTraining for staff on safe AI use and for developers on AI threats

Preventive controls such as guardrails reduce risk but never remove it. Detective controls and an owner who looks at the alerts are what make the control set work.

Layer preventive, detective and corrective controls, then report effectiveness to management.

Remember: a prompt or system prompt change can alter behaviour as much as a code change. Put prompts under version control and change management.

The full lesson continues in the interactive course on this page.

Sources and references

Unofficial practice course. Not affiliated with, endorsed by or sponsored by ISACA. AAISM, CISM and ISACA are trademarks of their owners; CISSP is a trademark of ISC2. All questions are original and are not real exam items. Check the current exam content outline on ISACA’s website before you book.

All courses · Cybersecurity and IT — Free courses, certificate $5 | AMAADOR ACADEMY