Specialists from Zhejiang University and Alibaba presented at ICML 2026 a new class of AI system attacks aimed not at hacking, but at slowing down requests to a state of useless processing. The method uses a genetic algorithm to create logical traps: problem conditions are intermixed, key assumptions are removed, prompting reasoning models to overthink. At MATH, the length of reasoning chains increased by 26.1 times. DeepSeek-R1, Qwen3-Thinking, GPT-o3 and Gemini 2.5 Flash were vulnerable. Queries created for small models are also effective against large systems, allowing you to prepare attacks at low cost.

The risk is particularly critical for AI agents in DeFi, trading, and smart contract auditing, where delays or logically faulted logic create direct operational and financial threats. The authors emphasize: the goal of the work is to fix a new surface of attacks, rather than demonstrate the possibility of mass break-ins. The study is based on previously identified patterns of "overthinking": paralytic analysis, unpredictable actions, and premature completion of tasks.