Get Started with AI Model Security

Table of Contents

Learn about the AI Model Security workflow and the step-by-step procedure to scan AI models.

Where Can I Use This?	What Do I Need?
Prisma AIRS (AI Model Security)	Prisma AIRS AI Model Security License

AI Model Security Workflow

AI Model Security operates on a two-tier hierarchical structure where each security group encompasses one or more security rules.

Source Type Binding: Model Security Groups are initially created and associated with a specific source type (such as, S3 buckets).
Rule Configuration: Populate these groups with relevant rules. AI Model Security provides intelligent suggestions and validation based on the selected source type, including default rule collections tailored to each source.
For instance, a "Verify Author" rule wouldn't be available for S3-based groups since S3 doesn't maintain author metadata.
Rules operate in either blocking or non-Blocking modes, streamlining the previous severity-based threshold system from AI Model Security.
Scanning Process: During model evaluation, the scan request specifies the applicable model security group to establish context. The system then processes two rule categories:
- Metadata Rules: Validate model metadata from the source platform
- Artifact Rules: Conduct comprehensive analysis of model files
Result Processing: Each rule produces a binary PASS or FAIL outcome. The final scan verdict aggregates all rule results and determines whether to apply blocking enforcement (for critical issues) or non-Blocking responses (for logging and alerts).
The AI Model Security delivers structured flexibility—the same model can undergo evaluation against multiple Model Security Groups (such as separate development and production configurations), each with distinct Source Types and rule sets, enabling context-appropriate security rules.