AutoRedTeamer: an automated red-teaming framework that continuously integrates new attacks

Researchers propose AutoRedTeamer, a fully automated end-to-end red-teaming framework for large language models that combines a multi-agent architecture with a memory-guided attack selection mechanism to continuously discover and integrate new attack vectors. The framework uses a dual-agent design: the red-team agent can generate and execute test cases based only on high-level risk categories, and the strategy-proposal agent autonomously discovers and implements new attacks by analyzing recent research; the modular design allows it to adapt to new threats while maintaining performance on existing attack vectors. The paper claims that on HarmBench, its attack success rate against Llama-3.1-70B is 20% higher than existing methods, with 46% lower compute cost, and that the diversity of generated test cases is comparable to human-curated benchmarks.