CART Framework Automates Adaptive Red Teaming for LLMs and Agents
September 24, 2026
CART is a closed-loop red teaming framework that uses a Challenger-Target-Judge architecture to adaptively generate probes based on previous failures. In evaluations across Frontier, JAH, and Agentic families, CART discovered higher average risk levels than static seed replay for all tested targets.
HOW THIS AFFECTS YOU
●
builderYou can use adaptive loops to find edge cases in tool-using agents that simple prompt replays will miss.
●
policyThis demonstrates that automated, adaptive testing is necessary to capture the evolving risk profiles of agentic systems.