Model Z.ai from Zhipu AI Matched Anthropic in Code Vulnerability Detection

Chinese model Z.ai from Zhipu AI has matched the performance of Anthropic's Mythos model in specific cybersecurity scenarios — particularly in detecting vulnerabilities in code, according to The Wall Street Journal citing researchers' assessments.

What the research showed

This is not about full parity, but about a specific narrow task: finding vulnerabilities in software code. In this scenario, according to WSJ data, Chinese model Z.ai was able to show results on par with Anthropic's powerful Mythos model.

• Z.ai model was developed by Chinese company Zhipu AI
• It matched Anthropic's Mythos model in code vulnerability detection tasks
• The assessment was made by researchers whose findings are cited by The Wall Street Journal
• In other cybersecurity tasks, the Chinese model still lags behind Anthropic and OpenAI models

Where China still lags the USA

Despite success in a narrow task, researchers note that overall Chinese models have not yet caught up with American ones in cybersecurity: outside of vulnerability detection, Z.ai falls short against both Anthropic and OpenAI systems. This means progress is selective rather than comprehensive — Chinese developers have managed to catch up with leaders in a specific class of tasks, but have not replicated their results across the full spectrum of cybersecurity scenarios.

Such dynamics are typical for the AI model race in recent years: companies choose narrow niches where they can quickly achieve parity with leaders, rather than competing across all directions at once. Cybersecurity is a particularly sensitive area, since models capable of finding vulnerabilities could theoretically be used both to protect systems and to attack them.

What this means

Even partial parity of Chinese models with Anthropic and OpenAI developments in cybersecurity tasks shows that the technological gap between the USA and China in applied AI scenarios continues to narrow — even if not yet across all directions at once.