“Perhaps most concerningly, it is unclear whether such behavior can ever be fully patched by LLM providers.”
CMU and Center for AI Safety researchers generate jailbreak suffixes automatically from open-source model weights, and the same strings work on ChatGPT, Bard and Claude. The vendors were warned before publication, so the published strings will stop working. The method won’t. Computer vision has failed to fix adversarial examples for 10 years. Alignment is a filter with a hole the shape of a string of junk characters.