Agentic red teaming can dramatically speed up security testing of AI systems by automating workflow construction, letting security teams focus on what vulnerabilities to test rather than how to implement each test.
This paper introduces an AI red teaming agent that automates adversarial testing of AI systems. Instead of manually building attack workflows over weeks, operators describe their testing goals in natural language, and the agent automatically selects attacks, applies transformations, and scores results—compressing the process from weeks to hours.