CISA and NSA Join Five Eyes Partners to Release Agentic AI Security Guidance

CISA and NSA, alongside Five Eyes partners from Australia, Canada, New Zealand, and the United Kingdom, published joint guidance on May 1 addressing the secure deployment of agentic AI systems in critical infrastructure and defense environments, CyberScoop reported. The guidance identifies five risk categories for autonomous AI agents: excessive system privileges, design and configuration flaws, unpredictable agent behavior, cascading failures across interconnected agent networks, and accountability gaps in opaque decision processes. The agencies recommend integrating agentic AI into existing cybersecurity frameworks using zero-trust and least-privilege principles, with each agent carrying a cryptographically verified identity and short-lived credentials. The document flags prompt injection as a persistent unresolved vulnerability and states that high-impact agent actions should require human approval.