What happened
On Nov 20, 2026 Palo Alto Networks announced the general availability of Prisma AIRS AI Supply Chain Security, extending its AI Model Security to agent skills. It scans open-source models and third-party agent skills inside CI/CD pipelines for malicious load-time/inference code, architectural backdoors, unsafe serialization (35+ model formats), embedded instructions that override agent behavior, credential access, and data-exfiltration paths. Models scan locally via SDK/CLI (proprietary models never leave the environment), skills can be scanned from GitHub, SDK, or ZIP upload, and 3M+ public Hugging Face models are pre-scanned. Policy is enforced as Security Groups with blocking in production vs monitoring in dev, with detection updated by Unit 42.
Why it matters
This is a GA launch of a genuinely new attack surface for AI security — the AI supply chain (models and agent skills), which conventional SCA/code scanners miss. It closes the gap where a single compromised model or skill artifact can spread into every agent that consumes it, and ties artifact verification into the existing Prisma AIRS platform.
Applicability
Defenders on Palo Alto Networks/Prisma AIRS should evaluate model and agent-skill scanning now; teams building on open-weight models or third-party agent skills should assess coverage. Skill scanning is currently North America-only and included at no extra cost for existing AI Model Security customers at launch.