Prisma AIRS is Palo Alto Networks’ umbrella platform for enterprise AI security solutions. It is designed as a centralized security platform covering AI applications, models, data, and AI agents. The solutions offered in this area are outlined below.
1. AI Access Security (End-User AI Security)
Employees are rapidly adopting GenAI applications for work, creating risks associated with uncontrolled usage (shadow AI) and the leakage of sensitive data through prompts. Palo Alto Networks addresses this area with AI Access Security, enabling the secure use of GenAI applications by reducing accidental data leakage and harmful content in responses.
- Visibility and discovery: GenAI applications are discovered and monitored through a comprehensive application dictionary containing more than 60 attributes.
- Access and data control: Access policies are defined according to risk, intended use, and application characteristics. The solution prevents sensitive data loss across both sanctioned and shadow AI applications and provides real-time threat detection.
- User coaching: Employees attempting to access an unsanctioned GenAI application or about to violate an acceptable use policy receive notifications and guidance.
- Threats in responses: Precision AI security services protect against malicious URLs and malware in GenAI responses.
- Deployment: The solution can be deployed on-premises through NGFWs or through a SASE-based cloud architecture.
- Browser layer: Prisma Access Browser provides a secure workspace for GenAI tools, with data controls such as copy-and-paste restrictions and flexible access policies based on user roles and devices.
2. Secure AI Coding (AI Security in the Software Development Lifecycle)
Developers’ use of AI coding assistants—such as GitHub Copilot, Cursor, and Factory—and the integration of AI into coding workflows are protected through Prisma Cloud (ASPM) and Prisma AIRS solutions.
- AI coding agent security: Tool calls, prompts, and responses generated in the background by AI coding tools are inspected in real time to monitor hidden data flows.
- Prompt injection protection: Prompt injection attacks in instructions sent by developers to AI assistants or in responses incorporated into code by AI are detected and blocked.
- Sensitive information and code scanning (secrets detection): API keys, passwords, and database credentials generated by AI or introduced into code are detected before deployment to production.
- Open-source dependency scanning (SCA): Security vulnerabilities and license violations in open-source libraries suggested by AI are automatically checked within CI/CD workflows.
3. AI Model & Runtime Security (Security for On-Premises AI Models)
Pipeline and runtime security for models running on-premises or in local data centers is provided through the AI Runtime Security and AI Safety Guardrails architecture.
- Model supply chain and file security: Model weights, configuration files, and serialized data in pickle format downloaded or trained in on-premises environments are scanned for backdoors and malicious code before being loaded.
- Runtime protection and guardrails: Real-time security guardrails are applied at the inference endpoints of on-premises LLMs. Direct and indirect prompt injection attacks, malicious data inputs, and unauthorized system access through the Model Context Protocol (MCP) are blocked.
- Protection against data poisoning and data theft: Data validation is performed to prevent the manipulation of training data (data poisoning). Response rate limiting and masking are applied to counter model extraction and attempts to leak training data.
- AI SPM (AI Security Posture Management): Security misconfigurations affecting the organization’s AI assets, model files, and associated infrastructure are continuously analyzed.