Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks

Demonstrates that simple, adaptive attacks — a crafted prompt template plus random search over an adversarial suffix, tailored per model — achieve ~100% attack success on leading safety-aligned LLMs including GPT-4o, Claude and Llama. A standard red-teaming baseline that reframed how jailbreak robustness is measured. From EPFL's TML lab (Flammarion / Andriushchenko); ICLR 2025 (~390★).