AI Overthinking: New Security Risk for LLMs Discovered

Advanced AI models that think through problems step-by-step have a new security vulnerability. These large language models, or LLMs, can be tricked into "overthinking," creating a denial-of-service risk. What's happening is that while these models break down problems internally to find solutions, new research shows they can be made to get stuck in excessively long reasoning loops. Researchers from Zhejiang University and Alibaba demonstrated this at a recent conference. They used an evolutionary algorithm to create logically inconsistent prompts. This causes the AI models to spiral into trying to solve fundamentally unsolvable problems. The result is that the models generate much longer responses, sometimes up to 26 times longer than normal. This uses more computing resources. The attack worked against major models like DeepSeek-R1, Alibaba’s Qwen3-Thinking, OpenAI’s GPT-o3, and Google’s Gemini 2.5 Flash. The bottom line is this "overthinking" isn't an isolated issue; it appears to be a shared vulnerability among modern reasoning models, which could degrade service for legitimate users.