What happened
The Prisma AIRS November 2026 release introduces a standardized three-action (Allow/Alert/Block) policy model across AI Security Profile detectors, Cloud IP-Tag Harvesting that maps public/hybrid Kubernetes cluster metadata into Dynamic Address Groups for real-time model-specific guardrails, Contextual Judge Reasoning for AI Red Teaming (upload domain context docs so the judge stops flagging intended behavior, and it learns from your overrides), and Dynamic Analysis for AI Skill Security — which loads a skill into an agent inside a sandbox and observes runtime behavior, with seven new dynamic rules covering outbound activity, credential access, exfiltration staging, process/shell execution, filesystem tampering, defense evasion, and persistence.
Why it matters
Adds runtime policy granularity (validate on live traffic before blocking), K8s-aware agent/inference guardrails that update as pods spin up, context-aware red-team judging that cuts false positives in domain-specific agents, and — most importantly — dynamic sandboxing of agent skills, closing the gap where static skill analysis misses exfiltration or destructive behavior.
Applicability
Palo Alto/Strata Cloud Manager shops running AI Runtime Security. Security teams governing LLM apps and agent skills should evaluate the dynamic skill analysis and the three-action policy model for agent and skill governance.