An agentic operating model is an AI architecture where the system does more than generate text: it can plan tasks, call tools or APIs, retrieve data, and complete multi-step actions with limited supervision. Instead of a single prompt-and-response flow, the model becomes part of an execution workflow that may touch files, tickets, email, cloud services, or internal databases.
This matters in cyber security because the risk is no longer only bad output, but bad action. If an attacker uses prompt injection, poisoned data, or a compromised integration, the agent may be pushed to expose information, approve changes, or trigger downstream operations. Defenders therefore treat the agent’s control plane as a security boundary and apply least privilege, scoped identities, approval steps, and runtime policy checks. Logs, traces, tool-call records, and provenance help prove what the agent saw, what it did, and whether it stayed within its mission.