An evaluation run is a controlled test used to observe how a model or agent behaves under repeatable conditions. In AI security, it usually means giving the system the same prompts, tools, websites, or policy constraints and then measuring what it does. Because the setup is consistent, researchers can compare behavior across versions, detect regressions, and identify unsafe actions before deployment.
Evaluation runs matter in cyber security because agentic systems can browse, click, fill forms, or follow instructions from untrusted content. A careful eval can reveal prompt injection risks, excessive permissions, data leakage, or unwanted tool use. Defenders use these runs to test guardrails, confirm that sensitive actions require approval, and verify that an agent behaves safely when it encounters hostile web pages or misleading content. In short, an evaluation run is one of the main ways to turn vague safety concerns into measurable security evidence.