What is agentic AI? Autonomous AI systems and their security risks, defined

Agentic AI describes systems built from one or more AI agents that use a model, usually a large language model (LLM), to interpret a goal, plan steps and take actions through tools, data sources and memory, with little or no continuous human direction. The term names a class of system and its degree of autonomy, not a single product or technique.

How the agencies define it

A detailed government definition comes from the joint guidance Careful adoption of agentic AI services, released in May 2026 by Australia's cyber agency with CISA, the NSA and the Canadian, New Zealand and UK cyber centres. It lists the parts of an LLM-based agentic system as the model, external tools, external data sources, memory and planning workflows, and says some such systems can spawn sub-agents of their own.

The same guidance separates it from generative AI: generative models create content for people, while agentic AI connects those models to software so they can reason, plan and act. The UK National Cyber Security Centre (NCSC), in advice published on 15 May 2026, frames the shift the same way, from producing predictions or content to accessing data, using tools and taking actions.

Why it matters to a CISO

Autonomy turns model errors into actions. The joint guidance says each tool, data source and memory store adds to an interconnected attack surface, and gives web search as a channel for indirect prompt injection. Its headline advice is never to give agentic AI broad or unrestricted access, and to keep it to low-risk, non-sensitive tasks.

The NCSC adds a governance test: be able to name who owns each agentic system, who approves its access, who monitors it, who reviews incidents and who can stop it. For risk registers, OWASP published its Top 10 for Agentic Applications on 9 December 2025, a peer-reviewed list built with more than 100 contributors. MITRE ATLAS now filters its attack matrix by an Agentic AI platform, covering components that plan, decide and execute multi-step actions.

Where it appears on cisonews.ai

The explainer how do you secure AI agents turns this guidance into controls. The regulation tracker records the Five Eyes agentic AI guidance and China's joint guidelines on AI agents. The incident log includes the first breach notification tied to an autonomous agent, filed with Spain's data protection authority.

Related terms

AI agent, Model Context Protocol (MCP), prompt injection and shadow AI.