AI models introduce a new, often invisible attack surface. They can contain:

• Embedded malicious code that runs at load or inference time
• Unsafe serialization formats (for example, Pickle) that allow arbitrary code execution
• Unverified or outdated dependencies that open exploit paths
• Poisoned data or hidden backdoors that change model behavior, such as silently bypassing fraud checks

Traditional security tools generally treat models as opaque black boxes, so these issues often go undetected. Prisma AIRS AI Model Security is designed to treat models as critical software assets and secure them from intake through deployment by:

• Deep model scanning (static and dynamic) of all artifacts: weights, binaries, configs, metadata, and dependencies
• Detecting deserialization threats, malicious payloads, backdoors, insecure libraries, and license violations
• Enforcing policies based on model source and usage (e.g., stricter rules for external models from Hugging Face vs. internal models in Amazon S3)
• Continuously monitoring deployed models with real-time dashboards, threat intelligence, and audit logs

The result is the ability to prevent model-borne threats before deployment, protect IP, and demonstrate governance and compliance without slowing AI development.

Prisma AIRS AI Model Security is built to plug into the tools and workflows you already use, so security becomes part of the pipeline rather than a separate, manual step.

Key integration points include:

• CI/CD pipelines: GitHub Actions, GitLab CI, Jenkins, Azure DevOps – models are scanned automatically at build and release time.
• Model registries: MLflow, Kubeflow, Vertex AI, and custom registries – models are validated as they are registered or promoted.
• Development environments: Jupyter, VS Code, PyCharm – developers and data scientists can trigger scans directly from their preferred tools.
• Deployment automation: Terraform, Ansible, Kubernetes – security checks can be embedded into infrastructure-as-code and deployment workflows.
• SIEM platforms: Cortex XSIAM and third-party aggregators – violations and alerts are centralized for security operations.

From an operations perspective, teams can:

• Activate scanning via APIs or SDKs for programmatic workflows
• Provision developer API keys from Strata Cloud Manager
• Use blocking or nonblocking rules to either stop risky models or just generate alerts

Performance-wise, Prisma AIRS is designed for high-scale environments, with API response times of 200 ms or less and a 99.9% uptime SLA. This helps you embed model security into everyday workflows without creating new bottlenecks.

Prisma AIRS AI Model Security is built to help you reimagine model security without sacrificing confidentiality.

To protect IP and sensitive data, Prisma AIRS:

• Performs local scanning so models do not need to leave your environment.
• Supports both on-premises and cloud deployment options, giving you control over where analysis runs.
• Adapts controls based on model origin – for example:
  • External models (e.g., from Hugging Face or partners): strict checks for malware, unsafe formats, license issues, and provenance.
  • Internal models (e.g., in Amazon S3 or local storage): focus on integrity, dependency validation, and version control without exporting artifacts.

For governance and compliance, Prisma AIRS provides:

• Audit-ready evidence – every scan verdict, rule change, and policy decision is logged.
• Exportable reports and evidence packages to support frameworks such as ISO and SOC 2, as well as industry-specific requirements.
• Real-time dashboards that show organizational risk trends and model posture over time.

Behind the scenes, Prisma AIRS is backed by a trusted database of over 1.9 million public AI models and continuously scans new releases from major repositories. This helps your security and compliance teams make informed decisions about which models to approve, reuse, or block, while keeping proprietary assets under your control.