DIA-7 - Artificial Intelligence Model Cybersecurity
Control for security across the AI/ML model lifecycle and protection of training data
Control Definition
Security and responsible use controls must be established for the development, deployment and operation of artificial intelligence models, assessing security, privacy, bias and regulatory compliance risks before adopting any model and requiring formal approval for the highest-risk ones. The exposure of personal or confidential data in interactions with AI models, both in inputs and outputs, must be prevented, and the use of organizational information to retrain third-party models restricted. AI-specific vulnerabilities must be incorporated into security assessment programs, and the supply chain of the models used must be managed.
Requirements
Requirement 1
We establish security and responsible use controls for the development, deployment and operation of AI models in the organization.
Requirement 2
We perform a risk assessment before adopting or deploying any AI model, whether developed internally, through third-party APIs or open-source models.
Requirement 3
In the risk assessment we analyze: security, privacy, bias, fairness and regulatory compliance risks.
Requirement 4
We require approval from the governance committee before deploying high-risk AI models.
Requirement 5
We prohibit including personal or confidential data in prompts sent to external AI models without prior assessment.
Requirement 6
We restrict the use of organizational data for retraining third-party models.
Requirement 7
We manage the AI model supply chain by assessing vendors, third-party APIs and open-source models.
Requirement 8
We implement output filtering controls to prevent data exfiltration through AI systems.
Requirement 9
We include AI-specific vulnerabilities (prompt injection, jailbreaking, model inversion, among others...) in security assessment programs.
Requirement 10
We ensure data isn't used to retrain the AI.