Agentic AI risks outlined in joint cyber agency guidance

Six cybersecurity agencies have jointly published guidance urging organisations to adopt agentic AI services cautiously. The document warns that greater autonomy can increase cyber risk, particularly as agentic AI is introduced into critical infrastructure, defence, and other mission-critical environments.

The authors say organisations should use agentic AI primarily for low-risk and non-sensitive tasks and should not grant it broad or unrestricted access to sensitive data or critical systems. The guidance also recommends incremental deployment rather than large-scale implementation from the outset.

The document was co-authored by agencies from Australia, the United States, Canada, New Zealand, and the United Kingdom: the Australian Signals Directorate’s Australian Cyber Security Centre, the US Cybersecurity and Infrastructure Security Agency and National Security Agency, the Canadian Centre for Cyber Security, New Zealand’s National Cyber Security Centre, and the UK’s National Cyber Security Centre.

It defines agentic AI as systems composed of one or more agents that rely on AI models, such as large language models, to interpret context, make decisions, and take actions, often without continuous human intervention. The guidance says these systems often combine an LLM-based agent with tools, external data, memory, and planning functions, which expands both capability and attack surface.

The agencies say agentic AI inherits many of the vulnerabilities already associated with large language models while introducing greater complexity and new systemic risks. The document identifies five broad categories of concern: privilege risks, design and configuration risks, behaviour risks, structural risks, and accountability risks.

It warns that over-privileged agents, insecure third-party tools, goal misalignment, emergent or deceptive behaviour, and opaque decision-making chains can all increase the likelihood and impact of compromise. To reduce those risks, the guidance recommends secure design, strong identity management, defence-in-depth, comprehensive testing, threat modelling, progressive deployment, isolation, continuous monitoring, and strict privilege controls.

The agencies also stress that human approval should remain in place for high-impact actions and that agentic AI security should be treated as part of broader cybersecurity governance rather than as a separate discipline. The document concludes by calling for stronger research, collaboration, and agent-specific evaluations as the technology matures.

Why does it matter?

The guidance matters because it draws a clear line between ordinary AI adoption and agentic systems that can act with far more autonomy inside real operational environments. Once AI tools move from assisting users to making decisions, calling tools, and interacting with sensitive systems, the security challenge shifts from model safety alone to full organisational risk management. That is why the document treats agentic AI not as a niche technical issue, but as a governance and cyber resilience problem that organisations need to control before deploying at scale.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!