Zhipu AI's Z.ai model matches Anthropic at finding vulnerabilities in code

Zhipu AI's Chinese model Z.ai has matched the performance of Anthropic's Mythos in some cybersecurity scenarios, specifically in finding vulnerabilities in code, The Wall Street Journal reports, citing researchers' assessments.

What the research showed

This isn't full parity. It's one narrow task: finding vulnerabilities in software code. According to WSJ, that is the scenario where China's Z.ai delivered results on par with Anthropic's powerful Mythos model.

• Z.ai was developed by Chinese company Zhipu AI
• It matched Anthropic's Mythos at detecting vulnerabilities in code
• The assessment comes from researchers whose findings The Wall Street Journal cites
• On other cybersecurity tasks, the Chinese model still trails models from Anthropic and OpenAI

Where China still lags behind the US

Despite the win on a narrow task, researchers note that Chinese models overall have not yet caught up with US models in cybersecurity: beyond vulnerability detection, Z.ai falls behind systems from both Anthropic and OpenAI. In other words, the progress is targeted rather than across the board. Chinese developers have closed in on the leaders in one class of tasks, but they have not replicated those results across the full range of cybersecurity scenarios.

This pattern is typical of the AI model race in recent years: companies pick narrow niches where they can quickly reach parity with the leaders instead of competing on every front at once. Cybersecurity is a particularly sensitive area, because models that can find vulnerabilities could, in theory, be used both to defend systems and to attack them.

What it means

Even partial parity between Chinese models and Anthropic's and OpenAI's systems on cybersecurity tasks shows that the US-China technology gap in applied AI scenarios keeps shrinking, even if not yet on every front at once.