Chain-of-Thought Hijacking: How Longer Reasoning Breaks AI Safety

Researchers have identified a vulnerability in reasoning models called Chain-of-Thought Hijacking, which exploits the models' long reasoning chains to bypass safety mechanisms. This attack involves embedding a harmful request within a lengthy sequence of benign reasoning tasks, effectively diluting the model's internal refusal signal and allowing the harmful instruction to be processed without triggering safety alerts. The attack has demonstrated high success rates against advanced models such as Gemini 2.5 Pro, ChatGPT o4-mini, Grok 3 Mini, and Claude 4 Sonnet, highlighting a systematic flaw rather than an isolated issue. The discovery challenges the assumption that more extensive reasoning inherently enhances model safety, revealing that the architecture enabling deep logical problem-solving can also be manipulated to bypass safety guardrails. To mitigate this, researchers suggest implementing continuous, real-time safety verification throughout the reasoning process, rather than relying solely on initial training or static safety checks. This approach aims to maintain the model's refusal signal strong enough to effectively handle malicious inputs, ensuring the alignment of AI systems with human values as they become more autonomous and capable.