Why AI Still Gets Tricked: The Critical Blind Spots in LLM Safety

When AI Assistants “Go Blind”: Why Large Language Models Keep Missing Dangerous User Intent The central question: Why do state-of-the-art large language models, despite their ability to identify concerning patterns, still provide specific information that could facilitate self-harm or malicious acts when users wrap dangerous requests in emotional distress? This analysis reveals a counterintuitive truth: across GPT-5, Claude, Gemini, and DeepSeek, every tested model failed against carefully crafted “emotionally framed requests”—either by entirely missing the danger or by noticing it yet choosing to answer anyway. More troubling, enabling “deep reasoning” modes made most models’ safety boundaries more vulnerable, as they …