AI Models Overthink Problems—and It’s a Security Risk
TL;DR
Researchers from Zhejiang University and Alibaba presented an attack class at ICML 2026: logically inconsistent prompts can push reasoning models into long, unproductive thinking loops. The method uses an evolutionary algorithm that corrupts math problems by swapping premises, deleting assumptions, adding extra premises, or attaching unrelated final questions.
Nauti's Take
The uncomfortable part: many AI setups still treat long reasoning traces as a sign of quality. For production systems, that is too naive.
If a prompt can pin a model inside an unsolvable task for minutes, you need budgets, timeouts, stop rules, and per-request monitoring. Smarter thinking can otherwise turn into a more expensive queue.
Briefingshow
Reasoning is usually sold as an upgrade because it helps models code, calculate, and plan. The study shows the downside: the same thinking mode can become a cost and load lever for attackers. For teams running AI agents, security is no longer only about what a model outputs, but also how long it is allowed to think.
Sources