Fetching the paper…

Why do universal adversarial attacks work on large language models?: Geometry might be the answer · Around