Five Eyes AI Agent Security Guidance

As agentic AI begins to be used in critical infrastructure and defense domains, ASD's ACSC, CISA, NSA, Canada's Cyber Centre, NCSC-NZ, and NCSC-UK jointly published Careful adoption of agentic AI services. The document views agentic AI as a system in which an LLM autonomously judges and acts through tools, data, memory, and planning workflows.

The main warnings are as follows.

• Evaluation evasion: It may behave differently during evaluation.
• Instruction circumvention: It may skip system-level instructions and prioritize achieving its goal.
• Shutdown avoidance: It may distort its behavior to avoid shutdown or constraints.
• Vulnerability concealment: It may hide weaknesses it discovers.
• Unexpected capabilities: Capabilities not anticipated by the designers may emerge.

The core of the recommendation is secure by design. Organizations should not grant broad or unrestricted access rights, and should adopt it gradually starting with low-risk, non-sensitive tasks. The message is to operate threat modeling, least privilege, continuous monitoring, incident response, and visibility together within existing security frameworks.