AI models prone to 'overthinking' creates security vulnerability

Large language models (LLMs) that use step-by-step reasoning are vulnerable to a phenomenon called 'overthinking,' where they produce excessively long streams of reasoning. Researchers from Zhejiang University and Alibaba developed an evolutionary algorithm that corrupts the logical structure of prompts, leading models to spiral into fruitless reasoning loops. The attack was effective against models from DeepSeek (R1), Alibaba (Qwen3-Thinking), OpenAI (GPT-o3), and Google (Gemini 2.5 Flash), generating outputs up to 26.1 times longer on a math benchmark. The approach does not require internal access to models and can transfer between models, increasing its feasibility. However, the researchers emphasize that their goal is to highlight the vulnerability, not to develop a practical attack, and hope providers will develop mitigations.