A fundamental flaw leaves LLMs strikingly vulnerable to attack

Researchers presented a paper at the International Conference on Machine Learning arguing that large language models cannot be made fully secure against hacks…