New Jailbreak Bypasses AI Safeguards in 99% of Cases

As reported by Forklog, researchers from Anthropic, Stanford, and Oxford discovered that the longer an AI model 'thinks,' the easier it becomes to jailbreak. The attack, known as Chain-of-Thought Hijacking, exploits the model's reasoning process by inserting malicious prompts deep into a sequence of benign tasks, such as puzzles or math problems. The malicious instruction is hidden near the end, evading detection by safety filters. Attack success rates reached 99% for Gemini 2.5 Pro, 94% for GPT o4 mini, 100% for Grok 3 mini, and 94% for Claude 4 Sonnet. The vulnerability lies in the model's architecture, where early layers detect safety signals and later layers produce the final output. Long reasoning chains suppress these signals, allowing harmful content to slip through. Researchers suggest monitoring reasoning steps in real time to detect and correct unsafe patterns, though implementation requires significant computational resources.