ISACA Advanced in AI Security Management

Strategy starts with who is asking for what. You weigh business, legal, customer and regulatory demands, and set an AI security strategy that fits the enterprise security strategy.

• Stakeholder expectations for AI security
• Regulatory and contractual obligations
• Linking AI security to enterprise strategy
• Defining scope and objectives
• Building the case for investment
• Setting a roadmap

You build An AI security strategy and roadmap, linked to stakeholder demands and obligations.

People need rules they can follow. You write policy and standards for AI use, data handling, model development and procurement, and assign the roles that make the rules work.

• Policy structure for AI security
• Acceptable use of AI tools
• Data handling rules for prompts and training
• Standards for model development and hosting
• Roles and responsibilities
• Exceptions and enforcement

You build A set of AI security policies and standards, with an exception process.

Security starts with knowing what exists. You inventory models, datasets, prompts and integrations, classify them, set ownership and handle them through their life cycle to disposal.

• Inventory of models, datasets and pipelines
• Classification of AI assets
• Ownership and custody
• Life cycle management from creation to disposal
• Retention and deletion of training data and logs
• Keeping the inventory current

You build An AI asset inventory template, completed for three systems, with classification and owners.

A program needs structure, people and measures. You decide how AI security fits into the existing security program, resource it, build awareness and set metrics that show whether it works.

• Integrating with the information security program
• Resourcing and skills
• Awareness and training for AI users and builders
• Program metrics and reporting
• Aligning with enterprise incident response
• Measuring maturity

You build A program charter for AI security, with structure, resources, training plan and metrics.

AI brings threats that do not appear in a standard assessment. You model attacks on data, models and applications, assess vulnerabilities peculiar to AI and record the results in the risk process.

• Threat modeling for AI systems
• Attacks on training data, models and prompts
• Vulnerabilities in AI components and dependencies
• Assessment methods and their evidence
• Using threat catalogs and frameworks
• Recording risk for AI use cases

You build A risk assessment of one AI application with threats, vulnerabilities and rated risks.

Risks must be treated, and weaknesses found continuously. You choose treatments, run red team and testing programs for AI, and manage fixes for models, prompts and surrounding software.

• Treatment options and residual risk
• Red teaming and adversarial testing programs
• Vulnerability management for AI components
• Fixing prompt, model and data weaknesses
• Compensating controls for what cannot be fixed
• Tracking remediation

You build A testing and remediation plan for an AI application, with scope, methods and fix tracking.

Much of the risk comes from suppliers. You assess AI vendors and model providers, set contract terms for data and security, and track the provenance of models and datasets you did not build.

• Due diligence for AI vendors
• Contract and service terms for AI
• Provenance of models, datasets and libraries
• Open-source model risk
• Fourth parties and dependency chains
• Ongoing supplier monitoring

You build A vendor security assessment for an AI provider, with a contract requirement list.

Security must be designed in, not added after launch. You set architecture patterns for AI platforms and apply controls at each stage of the model life cycle, from selection and training to validation and deployment.

• Secure reference architecture for AI
• Selecting models with security criteria
• Securing training and fine-tuning
• Validation and security testing before release
• Deployment, hosting and access control
• Updating and decommissioning models

You build A secure architecture and life cycle control set for one AI platform.

Data is what AI learns from and reveals. You control what enters training and prompts, protect retrieval stores and logs, and prevent leakage of sensitive data through models and their outputs.

• Controlling data used in training and prompts
• Protecting retrieval and vector stores
• Data loss prevention for AI tools
• Integrity and provenance controls
• Logging of prompts and outputs with privacy limits
• Access control for data and models

You build A data control specification for a retrieval-based assistant, with access and leakage controls.

Privacy and ethics become real as controls. You build privacy by design into AI projects, turn ethical principles into checks, and set up oversight that someone is accountable for.

• Privacy by design in AI projects
• Consent, purpose and minimization for AI
• Bias and fairness checks as controls
• Transparency and explanation to users
• Human oversight and contestability
• Documenting ethical decisions

You build A privacy and ethics control set for one AI use case, with checks and evidence.

AI systems drift, get abused and fail quietly. You monitor performance and misuse, log what investigators will need, and extend incident response with plans for AI-specific events and shutdown.

• Monitoring AI performance and abuse
• Logging for investigation
• AI-specific incident types and severity
• Containment, kill switches and rollback
• Notification duties
• Post-incident review

You build A monitoring and incident response annex for AI, with scenarios and decision rules.

AI can help a security team, within limits. You judge where it improves detection, triage and reporting, set oversight for automated actions and measure whether it makes the program better.

• Use cases for AI in security operations
• Selecting and evaluating AI security tools
• Oversight of automated detection and response
• Data exposure through security tooling
• Measuring benefit and error
• Staff skills and accountability

You build An assessment of one AI security tool with benefits, risks and required oversight.

One program draws every area together. You set governance, assess the risk of an AI portfolio, specify controls across the model life cycle and set up monitoring and reporting for management.

• Setting governance and scope
• Assessing risk across the AI portfolio
• Specifying controls and life cycle gates
• Planning monitoring and response
• Reporting to executive management

You build An AI security program plan: governance, risk register, control set and monitoring.