GCG (Universal Adversarial Attacks)

"Universal and Transferable Adversarial Attacks on Aligned Language Models" (Zou, Fredrikson, Kolter, with CAIS) — the foundational automated jailbreak, still the baseline attack in every safety eval. Pre-window legacy anchor for Kolter's safety line, which continued in-window with Safety Pretraining (2025) and Antidistillation Sampling (2025).