This control requires organizations to create, document, approve, communicate, apply, evaluate, and maintain formal audit and assurance policies and procedures for their AI systems. These policies should be reviewed and updated at least annually or whenever there are significant changes (such as deploying a new AI model, using new data sources, or responding to new regulations). Examples in AI: •If your organization uses an AI model for hiring, your audit policy should require regular checks for bias in model recommendations and ensure compliance with employment laws. •For a healthcare AI system, your procedures might include periodic reviews of model accuracy, data privacy, and explainability to meet regulatory requirements. •If you retrain your AI model with new data, your policy should require a review of the data source for quality and compliance before deployment.

• Are there documented audit and assurance policies specifically covering AI systems? • Are these policies approved by appropriate management or governance bodies? • Are the policies communicated to all relevant stakeholders (e.g., data scientists, engineers, compliance teams)? • Are the policies and procedures actually applied in practice for all AI projects? • Is there a process for evaluating the effectiveness of these policies? • Are the policies reviewed and updated at least annually or after significant changes? • Is there clear ownership assigned for maintaining and updating these policies? • Are audit logs and documentation maintained for all AI systems?

THREAT

Control/Measure:

This control requires organizations to have their AI systems and related processes reviewed by an independent party—someone not involved in the development, deployment, or daily operation of those systems—at least once a year. The assessment should follow recognized standards such as ISO 42001, NIST AI RMF, or other relevant frameworks. Examples in AI: •If your company uses an AI model for credit scoring, an external auditor might review your data handling, model validation, and fairness checks to ensure compliance with financial regulations. •For a healthcare AI system, an independent assessment could verify that patient data is handled securely and that the model’s predictions are explainable and accurate. •If you deploy a generative AI chatbot, an external review might evaluate how you monitor for harmful outputs and how you handle user data privacy.

•Have all critical AI systems undergone an independent audit in the past 12 months? •Are the auditors or assessors independent from the teams that develop, deploy, or operate the AI systems? •Are recognized standards (e.g., ISO 42001, NIST AI RMF) used as the basis for the assessment? •Are findings from independent assessments documented and tracked to resolution? •Are updates to threat models and controls made based on assessment results? •Is there a process for selecting qualified external assessors with AI and security expertise? •Are high-risk AI applications (e.g., those impacting safety or critical decisions) subject to additional external testing, such as red teaming or adversarial testing?

THREAT

Control/Measure

This control requires organizations to conduct independent audit and assurance assessments not just on a fixed schedule, but also in response to significant changes or new/emerging risks in their AI systems. These assessments should be guided by risk-based plans and policies, meaning you prioritize audits based on where the greatest risks exist. Examples in AI: •If you deploy a new AI model in a critical business process (e.g., fraud detection), you should trigger an independent assessment to ensure the new model doesn’t introduce unacceptable risks. •If a major vulnerability is discovered in a machine learning library you use, a risk-based assessment should be performed to check if your AI systems are affected. •If regulations change (e.g., new AI laws or privacy requirements), an independent audit should be conducted to ensure compliance. •If your AI system is exposed to new data sources or integrated with external APIs, a risk-based review should be triggered to assess potential threats.

•Are independent audits triggered by significant changes in AI systems (e.g., new model deployment, major updates, new data sources)? •Is there a documented risk-based plan or policy that defines when additional assessments are required? •Are emerging risks (e.g., new vulnerabilities, regulatory changes, or threat intelligence) regularly monitored and used to inform audit planning? •Are independent assessments conducted in response to incidents or near-misses involving AI systems? •Are the results of risk-based assessments documented and used to update risk registers and controls? •Is there a process for communicating the need for ad-hoc or event-driven audits to relevant stakeholders? •Are lessons learned from risk-based assessments incorporated into future risk management and audit planning?

THREAT

Control/Measure

This control requires organizations to verify that their AI systems and related processes comply with all relevant standards, regulations, legal, contractual, and statutory requirements during audits. This means you must ensure your AI activities are not only technically sound but also legally and ethically compliant. Examples in AI: •If your AI system processes personal data, you must verify compliance with data protection laws such as GDPR, CCPA, or HIPAA during audits. •For AI models used in financial services, you need to check compliance with sector-specific regulations (e.g., anti-money laundering, fair lending laws). •If you use third-party AI components or data, you must ensure contractual obligations (like data usage restrictions or licensing terms) are being followed. •If your organization is subject to AI-specific regulations (such as the EU AI Act), you must verify that your AI systems meet those requirements.

•Are all relevant standards, regulations, and legal requirements for AI systems identified and documented? •Is there a process to regularly review and update the list of applicable requirements as laws and standards evolve? •Are audits conducted to verify compliance with data privacy, security, and sector-specific regulations? •Are contractual and licensing obligations for third-party AI components and data sources reviewed during audits? •Are audit findings related to non-compliance documented and remediated in a timely manner? •Is there a process for training staff on relevant legal and regulatory requirements for AI? •Are compliance checks integrated into the AI development and deployment lifecycle?

THREAT

Control/Measure

This control requires organizations to verify that their AI systems and related processes comply with all relevant standards, regulations, legal, contractual, and statutory requirements during audits. This means you must ensure your AI activities are not only technically sound but also legally and ethically compliant. Examples in AI: •If your AI system processes personal data, you must verify compliance with data protection laws such as GDPR, CCPA, or HIPAA during audits. •For AI models used in financial services, you need to check compliance with sector-specific regulations (e.g., anti-money laundering, fair lending laws). •If you use third-party AI components or data, you must ensure contractual obligations (like data usage restrictions or licensing terms) are being followed. •If your organization is subject to AI-specific regulations (such as the EU AI Act), you must verify that your AI systems meet those requirements.

•Are all relevant standards, regulations, and legal requirements for AI systems identified and documented? •Is there a process to regularly review and update the list of applicable requirements as laws and standards evolve? •Are audits conducted to verify compliance with data privacy, security, and sector-specific regulations? •Are contractual and licensing obligations for third-party AI components and data sources reviewed during audits? •Are audit findings related to non-compliance documented and remediated in a timely manner? •Is there a process for training staff on relevant legal and regulatory requirements for AI? •Are compliance checks integrated into the AI development and deployment lifecycle?

THREAT

Control/Measure