Beijing-based Zhipu AI’s GLM-5.2 system is just as good at finding gaps and bugs as Anthropic’s Mythos, according to researchers. It’s still not as good as tools offered by the Claude maker or OpenAI at other tasks, though. (The Wall Street Journal)

Talking point: AI models that can find security vulnerabilities—and then exploit or patch them, depending on who’s controlling the system—have alarmed policymakers and business leaders, and prompted the U.S. to install a de facto licensing system. Both Anthropic and OpenAI have limited access to their newest, most advanced systems, the latter at the behest of the Trump administration. But Zhipu’s new model, like many Chinese AI tools, is open weight, meaning anyone can download and use it. Earlier this month, researchers at the University of Toronto showed that it’s possible to use such freely available tools to create “adaptive computer worms,” malware that will spread across networks on its own.