Red-teaming is the practice of deliberately trying to break, manipulate, or misuse an AI system before it reaches real users. The term comes from military exercises where one team (the "red team") plays the adversary to expose weaknesses. In AI, red-teamers craft tricky inputs designed to make a model behave unsafely, leak information, or bypass its guidelines.

For agentic AI systems — models that can take actions in the world like browsing the web, writing code, or managing files — red-teaming becomes especially critical. A basic chatbot that says something wrong is annoying; an agent that takes a wrong action can cause real damage. Red-teaming agents means testing not just what they say, but what they do across long, multi-step tasks where harmful behavior might only emerge after several interactions.