China's Zhipu AI releases GLM-5.2, claims parity with Mythos on cybersecurity benchmarks

Zhipu AI released its open-weight GLM-5.2 model on June 28, 2026, with researchers claiming it matches Anthropic's Mythos on certain bug-finding and cybersecurity scenarios. The model represents a significant narrowing of the capability gap between Chinese and US AI systems in specialized security tasks.

Performance and capabilities

According to researchers, GLM-5.2 achieves performance comparable to Mythos on specific cybersecurity benchmarks focused on vulnerability detection and bug-finding. However, the model still lags behind Anthropic and OpenAI models on general-purpose tasks. Specific benchmark scores and pricing details have not been disclosed.

The model is released with open weights, allowing anyone to download and run it on readily available hardware without the restrictions typically imposed on frontier models.

National security implications

The development has drawn attention from US government officials who have worked to restrict China's access to advanced AI models like Anthropic's Mythos and Fable, as well as the hardware required to train and deploy them. The Trump administration has classified advanced AI models capable of identifying security vulnerabilities as national security threats.

OpenAI recently unveiled GPT-5.6, which has similarly raised concerns about potential misuse, leading to limited access controls.

Open-weight distribution concerns

Unlike restricted commercial models, GLM-5.2's open-weight release allows power users deep access and flexibility but also enables potential abuse by malicious actors who can run the model with minimal oversight. This distribution model eliminates many of the safeguards typically implemented by AI companies for sensitive capabilities.

What this means

GLM-5.2's claimed parity with Mythos on cybersecurity tasks, if verified independently, indicates China has successfully developed competitive AI capabilities in specialized domains despite US export restrictions. The open-weight release bypasses access controls that governments and companies use to prevent misuse of powerful security-focused AI systems. This creates a new dynamic where cutting-edge vulnerability detection capabilities become widely accessible, potentially benefiting both security researchers and threat actors equally.