Giorgio Piras

New attack method bypasses language model safety features

Researchers have developed a new method called Controlled Latent-space Evasion (CLE) to bypass safety mechanisms in language models. This technique reframes refusal suppression as an attack on the model's internal repre…