How Prompt Attacks Exploit GenAI and How to Fight Back

This report from Palo Alto Networks outlines a comprehensive taxonomy and defense framework for adversarial prompt attacks on GenAI applications and AI agents, categorizing impacts (goal hijacking, guardrail bypass, information leakage, infrastructure attacks) and techniques (prompt engineering, social engineering, obfuscation, knowledge poisoning). It underscores the urgency of GenAI security by noting high attack success rates (up to 88% on some models), explores risks introduced by RAG and agentic workflows, and recommends multi-layered mitigations such as adaptive prompt guardrails, sensitive data and payload scanning, and stronger access controls—alongside references to supporting security solutions.