New Technique Exposes LLM Vulnerabilities in Safety Measures

Researchers have introduced Incremental Completion Decomposition (ICD), a new strategy aimed at breaking the safety mechanisms of Large Language Models (LLMs). This approach uses a trajectory-based method to exploit safety vulnerabilities, allowing for a sequence of single-word responses that lead to harmful outputs. Evaluations across multiple model families show that ICD significantly outperforms traditional attack methods in terms of Attack Success Rate (ASR).