- Zhipu AI’s open-weight model GLM-5.2 has matched Anthropic’s Mythos in some cybersecurity bug-finding benchmarks, according to security researchers — and ranks among the 10 most-used AI models globally, while being freely downloadable and modifiable by anyone, including hackers.
- The Trump administration made the gap worse: OpenAI limited access to GPT-5.6 over security concerns, and Anthropic’s general-use “Fable” model was shut down for over two weeks — cutting off the NSA, which had been testing it and found it “impressive” — before partial access to the related “Mythos 5” was restored Friday.
- Critics say the White House approach is self-defeating: “Banning Fable while selling chips China needs to develop its own version is a gift to China,” said Saif Khan of the Institute for Progress — while the administration has been permissive about Chinese open-weight models like DeepSeek and Zhipu proliferating inside US businesses.
- China’s 360 Security Technology also released a new bug-finding tool called Tulongfeng on Wednesday, claiming parity with Mythos, as the overall capability gap between top US and Chinese AI models has “narrowed significantly” and Chinese AI usage has surged as businesses seek to cut costs.
What Happened?
Chinese AI firm Zhipu AI released GLM-5.2, an open-weight model that security researchers say can match Anthropic’s Mythos in cybersecurity bug-finding scenarios, besting Anthropic’s Claude Opus 4.8 in some Semgrep benchmarks. GLM-5.2 is open-weight — anyone can download, modify, and run it without supervision, including bad actors. Separately, China’s 360 Security Technology released Tulongfeng, a bug-finding tool its CEO claimed is comparable to Mythos. Meanwhile, the Trump administration restricted OpenAI’s GPT-5.6 and shut down Anthropic’s Fable model entirely for two weeks — blocking the NSA from tools it had been testing. On Friday, the administration restored some access to Anthropic’s Mythos 5 for trusted entities.
Why It Matters?
AI models capable of finding software vulnerabilities are strategic assets in cyberwarfare — whoever finds bugs first can exploit or patch them. China’s GLM-5.2 being open-weight means it’s accessible to state-sponsored hackers with zero restrictions, while US labs face an administration that restricts their own customers. The NSA lost access to tools it found impressive; the intelligence community’s ability to harden US cyber defenses was interrupted. As Saif Khan of the Institute for Progress put it, the US “needs to maximize the use of Mythos and comparable models to harden its cyber defenses while it can.” Instead, critics say Washington is simultaneously ceding the capability edge and enabling Chinese competition through chip sales.
What’s Next?
The Trump administration says it is “tracking very closely” Chinese open-weight models, per Undersecretary of State Jacob Helberg — but no restrictions have materialized. The Pentagon recently signed a deal with Reflection AI, one of the few domestic open-weight developers, for classified use. Microsoft is weighing whether to offer Chinese models on its platforms, which would fundamentally alter the competitive landscape among big tech AI distributors. The fundamental tension — between controlling powerful models and maintaining US AI leadership — is now a live policy crisis with no obvious resolution in sight.
Source: The Wall Street Journal
