Back to News
NewsAi

Researchers Learn to Slow Down AI Models with Logical Traps

Researchers have developed a new method to slow down AI models using logical traps, presenting a potential denial-of-service attack vector.

Mercation Admin
July 8, 2026 2 min read
Researchers Learn to Slow Down AI Models with Logical Traps

Researchers from Zhejiang University and Alibaba unveiled a novel attack strategy targeting AI systems at the ICML 2026 conference in Seoul. This approach aims to incapacitate AI models by significantly delaying their response times, rather than compromising their data security.

Mechanics of the Attack

Distinct from traditional large language models (LLMs), reasoning models tackle tasks through a series of sequential steps, making them ideal for complex analyses. However, when faced with incomplete or contradictory information, these models often engage in extensive reasoning, leading to longer processing times and greater resource consumption. This behavior creates vulnerabilities that can be exploited for denial-of-service (DoS) attacks.

The researchers engineered a method that intentionally triggers this excessive reasoning. By employing a genetic algorithm, they modified task parameters, removing essential premises while introducing irrelevant ones, and selected scenarios that prompted the longest response times.

The results were striking, with reasoning length on the MATH benchmark increasing by 26.1 times. This technique proved more effective than previous methods, affecting various models including DeepSeek-R1, Qwen3-Thinking, GPT-o3, and Gemini 2.5 Flash.

Moreover, the study revealed that queries optimized for smaller models could also disrupt larger commercial systems, indicating a low-cost attack vector for closed service environments.

Wei Cao, one of the researchers, stated, “Our goal is not to demonstrate that large-scale attacks are possible with minimal costs, but to highlight that this attack surface exists.”

Implications of the Findings

As reasoning models gain traction in various AI applications, including trading algorithms and smart contract audits, the risks associated with logical failures become more pronounced. In decentralized finance (DeFi), for instance, AI systems manage real assets autonomously, making them susceptible to operational risks stemming from erroneous logic.

This research builds on existing knowledge regarding the tendency of reasoning models to overthink. Earlier studies identified patterns such as:

  • analysis paralysis — models continue reasoning indefinitely instead of concluding tasks;
  • unpredictable actions — they may attempt multiple actions simultaneously after encountering errors;
  • premature termination — models stop executing tasks without confirming their results.

The propensity for overthinking was found to negatively impact the models' effectiveness. Additionally, analysts have cautioned that the future of AI development by companies like OpenAI and Anthropic increasingly hinges on their access to computational resources, funding for data centers, and evolving regulatory landscapes.