AI Reasoning Flaw Creates New Security Risk: Overthinking

Advanced AI models, particularly large language models (LLMs) capable of step-by-step reasoning, present a new security vulnerability where attackers can induce "overthinking," significantly slowing down these systems. Unlike earlier models that provided direct answers, current LLMs generate an internal monologue to break down complex problems, enhancing their capabilities in areas like coding and mathematics. However, research presented at the International Conference on Machine Learning 2026 in Seoul by scientists from Zhejiang University and Alibaba reveals that logically inconsistent prompts can deliberately trigger excessive reasoning. This "evolutionary prompt attack" corrupts prompts, causing models to spiral into unproductive thought processes when faced with unsolvable problems. The resulting longer outputs increase server load and costs for AI providers, potentially degrading service for legitimate users. The attack proved effective against models from DeepSeek-R1, Alibaba's Qwen3-Thinking, OpenAI's GPT-o3, and Google's Gemini 2.5 Flash, producing outputs up to 26 times longer on math benchmarks. Researchers demonstrated that this overthinking is a shared vulnerability across modern reasoning models, not an isolated issue. The attack method involves using a genetic algorithm to mutate logical structures of problems, increasing output length and overthinking markers. This approach does not require internal model access, making it applicable to closed-source commercial services. While prompt generation can be resource-intensive, the researchers showed that prompts created by smaller models can still effectively trigger overthinking in larger target models, increasing the attack's feasibility. The researchers emphasize that their goal is to highlight this attack surface and encourage mitigation, rather than to deploy a practical denial-of-service attack.