Chinese AI closes the gap with Anthropic in cybersecurity

Open Chinese AI models are beginning to approach Anthropic's most powerful models in certain cybersecurity tasks.

According to The Wall Street Journal, GLM-5.2, a model released in mid-June by Zhipu AI with openly available weights, matched Anthropic's Mythos in some cybersecurity scenarios. The result is notable because U.S. authorities have allowed access to Mythos only for selected American companies, while Fable 5, its open-weight counterpart, has been blocked for foreign users.

The assessment comes from cybersecurity specialists at San Francisco-based Semgrep, who tested GLM-5.2 on their own tasks. According to the company, the model "outperformed" Anthropic's Claude Opus 4.8 in vulnerability detection and, with more advanced prompting, "could match" Mythos-class models, meaning Mythos 5 and Fable 5. The key advantage of the Chinese model is that it can be used not only through an API or chat interface, but also run locally on a user's own computer, while access to Anthropic's models remains restricted.

The comparison, however, applies only to a narrow set of tasks, not overall model capability. In benchmarks from independent research firm Artificial Analysis measuring general performance and intelligence, GLM-5.2 ranks fourth, behind OpenAI's GPT-5.5. In coding, it ranks sixth, also behind models including Google's Gemini 3.1 Pro. Even so, the model is among the ten most popular on OpenRouter, an aggregator developers use to access different AI models.

According to the WSJ, another Chinese tool, Tulongfeng from cybersecurity company 360 Security, is also comparable to Mythos-class models. The publication did not specify which underlying models power it.