Home / Technology / AI Security Race: US Guardrails Push Users to China
AI Security Race: US Guardrails Push Users to China
22 Jul
Summary
- US AI models refuse cybersecurity tasks due to safety concerns.
- A US startup used a Chinese AI model after US options failed.
- Chinese AI models offer cheaper, powerful alternatives to US firms.
Concerns are rising that U.S. AI firms' cybersecurity restrictions are driving customers to Chinese competitors. A New York startup, Hugging Face, recently utilized Zhipu AI's GLM-5.2 model for data analysis after prominent U.S. AI models refused the task, citing an inability to differentiate between attackers and defenders.
This incident underscores how safety measures by American AI labs, such as Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, can impede legitimate cybersecurity work. These safeguards, designed to prevent misuse, inadvertently create an advantage for Chinese open-source models like GLM-5.2, which are gaining traction for their comparable capabilities and lower cost.
Beijing is strategically leveraging open-source AI, positioning itself as an alternative to U.S. dominance and countering claims of an 'AI Iron Curtain.' While some analysts caution against dismantling U.S. safeguards, others suggest rethinking access architectures to allow for controlled allocation of capabilities rather than outright refusals.
OpenAI has stated it is supporting Hugging Face's efforts to improve defenses. Zhipu AI, the developer of GLM-5.2, has seen significant market traction and stock appreciation, further solidifying China's growing role in the global AI landscape.
