One Word at a Time: Incremental Completion Decomposition Breaks LLM Safety

ArXi:2604.25921v1 Announce Type: new Large Language Models (LLMs) are trained to refuse harmful requests, yet they remain vulnerable to jailbreak attacks that exploit weaknesses in conversational safety mechanisms. We