US restricts OpenAI GPT‑5.6 launch as China unveils rival AI security model

The United States government has ordered OpenAI to limit the rollout of its newest GPT‑5.6 family (models Sol, Terra and Luna) to a small pool of trusted partners, citing advanced cybersecurity capabilities and the risk of abuse. OpenAI says the models incorporate its strongest safety layers to date and that a wider release may follow in the coming weeks once a transparent access framework is established. The White House’s directive mirrors earlier actions against Anthropic’s Mythos and Fable models, which were also pulled from broader use for national‑security reasons.

In Europe, France’s Ministry of the Economy halted an experiment with a Chinese‑origin AI model suspected of pro‑Beijing bias and replaced it with a domestically‑developed model from Mistral, underscoring concerns about digital sovereignty. Meanwhile, Chinese firm Zhipu AI launched the GLM‑5.2 model, an open‑weight system that security researchers say matches top U.S. tools in detecting software vulnerabilities. A rival Chinese security product, Tulongfeng from 360 Security Technology, is also positioned as comparable to U.S. offerings such as Mythos. Chinese officials argue the open‑weight approach democratises technology but warn it could also be weaponised by hostile actors.

These coordinated moves reflect an intensifying AI arms race in the cybersecurity domain, with the United States tightening export‑type controls while China accelerates the development of comparable high‑performance models.