Double Agents: Exposing Security Blind Spots in GCP Vertex AI

Palo Alto Networks Unit 42 released multiple March 2026 research and briefings highlighting emergent operational risks from agentic AI and nation-state activity. Key findings include a “Double Agents” flaw in Google Cloud Vertex AI where overprivileged AI agents can be abused to escalate access and compromise cloud resources; multiple research pieces demonstrating prompt-injection and prompt-fuzzing techniques that bypass LLM guardrails and “AI judges”; broader analysis of agentic AI abuse vectors (retail fraud, AI-enabled malware, and security tradeoffs when granting agents excessive rights);