A natural-language scaffold with predefined tools let an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → reproduce multistage attack chains in hours, speeding defender red teamingDeliberately attacking your own AI system to find what makes it fail before someone else does.Full definition →. It also shows agents improvising around failed tools, relevant when scoping agent permissions.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
red teaming — Deliberately attacking your own AI system to find what makes it fail before someone else does.
