US government, allies publish guidance on how to safely deploy AI agents

The Five Eyes cybersecurity agencies released joint guidance urging organizations to treat agentic AI (autonomous, action-taking systems built on large language models) as a core cybersecurity concern. The guidance identifies five risk categories — excessive privileges, design/configuration flaws, unintended agent behavior, structural risks from interconnected agents, and accountability gaps — and recommends integrating agentic systems into existing cybersecurity frameworks (zero trust, defense-in-depth, least privilege), enforcing cryptographic identities and short-lived credentials, encrypting communications, requiring human approval for high-impact actions, and prioritizing resilience and risk containment while research and standards mature.