New Threats to Language Models: Breaking Security Filters

A study has revealed a new method for bypassing security filters in language models that may lead to the leakage of prohibited information.