Researchers identify fundamental security flaw in LLMs

A paper presented at the International Conference on Machine Learning argues that LLMs are inherently vulnerable to hacks due to a structural flaw. The researchers claim it is impossible to fully secure these models against adversarial attacks.