AI Models Overthink Problems—and It’s a Security Risk
TL;DR
New research from Zhejiang University and Alibaba shows that reasoning models can be pushed into deliberate overthinking with logically inconsistent prompts. The team used an evolutionary prompt attack that mutates premises and questions until models produce unusually long answers and hesitation markers. Models tested included DeepSeek-R1, Qwen3-Thinking, OpenAI GPT-o3 and Gemini 2.5 Flash. On one math benchmark, outputs grew up to 26.1 times longer.
Nauti's Take
This is a useful warning about an underrated class of AI risk: safety is not only about jailbreaks and data leaks. A model that thinks for too long can become economically and operationally fragile, much like an overloaded server.
The study does not prove a cheap mass attack, but it does expose a real lever. Reasoning models need stop rules, contradiction detection and strict compute budgets, or their intelligence turns into a cost trap.
Briefingshow
Reasoning is one of the main selling points of modern AI models, but it also widens the attack surface. If a model cannot stop early on unsolvable or contradictory tasks, an attacker can exploit compute rather than steal data. For providers, prompt filtering becomes a cost and availability security issue, not just a content safety layer.
Sources