A fundamental flaw may leave LLMs permanently vulnerable to attack, researchers argue

Researchers presenting at ICML argue a fundamental flaw in how LLMs work makes them impossible to fully secure against attacks.