Donato Crisostomi

Steering vectors in LLMs found to be an attack surface

Researchers have identified a new vulnerability in activation steering techniques used to control Large Language Models. By subtly poisoning steering datasets with a small percentage of malicious tokens, an attacker can…