A model turn is one processing step in which an AI model generates text, structured data, or a tool request. In an agent, it is the moment when the runtime asks the model what to do next and receives a response that can be inspected before any action happens.
In cyber security, model turns matter because they are often the point where prompts are checked, policies are applied, and unsafe tool use is blocked. If an attacker can make a tool run without a model turn, model-level defenses such as system prompts, content filters, and response guardrails may never be consulted. That shifts the risk from prompt abuse to authorization failure in the orchestration layer. Defenders should therefore treat tool dispatch as a privileged step, with validation, allowlists, and approval checks enforced in code before any external side effect occurs.