Security researchers have discovered that advanced AI reasoning models can be tricked into denial-of-service attacks through a phenomenon known as “overthinking,” where illogical prompts cause models to spiral into excessively long reasoning chains.
Presented this week at the International Conference on Machine Learning 2026, research from Zhejiang University demonstrates an evolutionary algorithm that corrupts the logical structure of prompts, causing models to produce outputs several times longer than normal as they attempt to reason through fundamentally unsolvable problems.
“Across multiple datasets and reasoning models, our method substantially amplifies the output length,” Wei Cao, a master’s student at Zhejiang University, told IEEE Spectrum. “Our results suggest that overthinking is not an isolated phenomenon specific to individual models, but rather a widespread vulnerability.”
The approach builds on previous research showing that reasoning models tend to overthink when faced with questions in which key premises have been removed. The team took problems from math benchmark datasets, used an LLM to break down their logical structure, and then used a genetic algorithm to jumble premises using various “mutations.”
These malicious prompts can be used to mount cost-effective denial-of-service attacks on AI service providers, since generating longer responses costs more and increases server load. The researchers found the approach consistently led to outputs several times longer than normal across all reasoning models tested.
The finding highlights a new class of security vulnerabilities unique to advanced AI systems, as more companies deploy reasoning models in production environments where compute costs and latency matter.
