Researchers have learned to slow down AI models with logical traps - ForkLog

# Researchers Learn to Slow Down AI Models with Logic Traps

Specialists from Zhejiang University and Alibaba presented a new class of attacks on AI systems at ICML 2026 in Seoul, writes IEEE Spectrum. Their goal is not to hack the model or gain access to data, but to make it process queries for so long that it becomes useless.

How the New Method Works

Reasoning models — unlike ordinary LLMs — break down a task into sequential steps before responding. They are increasingly used in systems that require complex multi-stage analysis.

When working with incomplete or contradictory data, such models tend to overthink — generating excessively long chains of reasoning. This increases query processing time and computational resource consumption. In automated systems, this opens a vector for DoS attacks.

The researchers developed a method that intentionally provokes this behavior. A genetic algorithm shuffles task conditions, removes key premises, and adds redundant ones. It then selects variants that trigger the longest possible response.

On the MATH benchmark, reasoning length increased 26.1 times. The method outperformed existing ways of such impact. DeepSeek-R1, Qwen3-Thinking, GPT-o3, and Gemini 2.5 Flash proved vulnerable.

The authors also found that queries created for one small model proved effective against other systems, including large commercial projects. This allows preparing attacks on closed services at low cost.

"Our goal is not to demonstrate that large-scale attacks are possible at minimal cost, but to document that this attack surface exists," wrote one of the researchers, Wei Cao, in a letter to IEEE Spectrum.

Why This Matters

Reasoning models are increasingly used in agentic AI systems, including trading bots, smart contract audit tools, and decentralized infrastructure.

In DeFi, AI-powered digital assistants manage real funds without human intervention. A logic failure — including one intentionally caused — creates operational risk.

The new work builds on a known feature of reasoning models — a tendency to overthink. In February 2025, a group of researchers analyzed 4,018 agent trajectories and identified recurring patterns of overthinking in models:

• analysis paralysis — the model continues reasoning instead of executing the task;
• unpredictable actions — after an error, it attempts to execute multiple actions simultaneously;
• premature termination — stops task execution without verifying the result.

Reasoning models proved more prone to overthinking. The stronger the effect, the lower the performance.

Recall that in early July 2026, analysts warned that the further development of OpenAI and Anthropic increasingly depends on the availability of computing power, data center funding, and regulatory decisions.