Attribute-Based Access Control for AI

ABAC provides dynamic, context-aware access control by evaluating attributes of users, resources, actions, and the environment. For AI systems, ABAC enables fine-grained policies that adapt to data sensitivity, project context, and compliance requirements.

ABAC Components

ABAC policies evaluate four categories of attributes:

ABAC Policies for AI

# Allow model training only on non-PII data
# unless user has PII-certified clearance
allow {
    input.action == "train"
    input.resource.sensitivity != "pii"
}

allow {
    input.action == "train"
    input.resource.sensitivity == "pii"
    input.subject.clearance == "pii-certified"
    input.environment.network == "secure-vpc"
}

AI-Specific ABAC Use Cases

• Data sensitivity gating: Only allow training on PII data when users have completed privacy training and are on the secure network
• Model deployment approval: Require models to pass fairness and safety evaluations before allowing deployment to production
• Cross-project data access: Allow data sharing between projects only when both projects have compatible compliance classifications
• Inference rate control: Adjust rate limits based on user tier, application priority, and current system load
• Geographic restrictions: Restrict access to models trained on region-specific data to users in that region

ABAC vs RBAC: When to Use Each