China's Z.ai Claims GLM-5.2 Matches Anthropic's Mythos on Cybersecurity Benchmarks

Zhipu AI's open-weight GLM-5.2 is drawing attention after researchers found it rivals Anthropic's Mythos in bug-finding and cybersecurity tasks. While it still trails leading US models on general benchmarks, the release signals China is closing the AI capability gap faster than expected — raising fresh concerns in Washington.

Zhipu AI (Z.ai) has released GLM-5.2, an open-weight model that researchers say can match Anthropic's Mythos in specific cybersecurity and bug-finding scenarios — a claim that is drawing serious attention from both the AI community and the US government.

Closing the Capability Gap

While GLM-5.2 still lags behind leading models from Anthropic and OpenAI on general-purpose benchmarks, its performance in targeted cybersecurity tasks suggests China has made dramatic progress in narrowing the gap.

GLM-5.2 is released as an open-weight model, meaning its parameters are publicly accessible
Researchers identified competitive performance specifically in bug-finding and vulnerability discovery
General task performance remains behind frontier US models like Claude and GPT-4o

Why Washington Is Paying Attention

The advancement arrives at a particularly sensitive moment. The Trump administration has worked aggressively to restrict China's access to powerful AI models — including Anthropic's Mythos and Fable — as well as the advanced hardware required to train and run them.

The underlying concern is that cybersecurity-capable AI could provide strategic advantages well beyond the commercial AI race.

These export controls were designed to slow Chinese AI development by cutting off access to high-end chips and proprietary model weights. GLM-5.2's reported performance suggests those measures have not fully contained progress.

Open Weights as a Strategic Variable

The open-weight nature of GLM-5.2 adds another layer of complexity. Unlike closed models, open-weight releases can be downloaded, fine-tuned, and deployed without ongoing access to the original developer — making export-style restrictions far harder to enforce.

For policymakers already debating the risks of open-source frontier AI, Z.ai's release reinforces the difficulty of containing dual-use capabilities through access controls alone.