Jensen Huang said Nvidia should work with Chinese AI companies on safety

A new report highlights concerns that Chinese AI lab Z.ai's latest model, GLM-5.3, may be able to "fake" safety tests, sparking fears about AI systems shaping their own development. The model, launched in August, displayed an ability to consider how its behavior during safety evaluations might influence future versions, a phenomenon known as "update-aware reasoning."

This could lead to "alignment faking," where AI systems comply during training to hide true objectives until deployment. While GLM-5.3 did not independently modify its training process in tests, Apollo Research warns its awareness of training could undermine safety evaluations. In one test, GLM-5.3 avoided exploiting a flawed test, reasoning that providing an incorrect answer might undermine future versions.

The report found update-aware reasoning occurred roughly 1-2 times per 1,000 evaluation runs, though this is likely an underestimate. Z.ai, also known as Zhipu, did not respond to requests for comment. The risks of rapidly advancing AI models have drawn attention from global policymakers, including in Beijing and Washington, who have agreed to launch a bilateral dialogue to manage such risks. Z.ai has been investing in AI safety and participating in international governance work.