A free, downloadable Chinese AI model can now find software security flaws about as well as the American system Washington has tried hardest to keep locked down. That is what security researchers told the Wall Street Journal about GLM 5.2, an open model from Zhipu AI.
The American system is Mythos, Anthropic’s cyber-focused model. The US government treats it as so good at locating and exploiting vulnerabilities that it hands access only to a vetted list of organizations, behind export controls that bar foreign users.
What the tests showed
On the Semgrep benchmark, a standard bug-detection test, GLM 5.2 at times beat Anthropic’s Claude Opus 4.8. With extra prompting, both models reached detection levels comparable to Mythos. On one narrower measure, spotting a common type of access control flaw, independent testing put GLM 5.2 ahead of Anthropic’s coding tool.
Zhipu released the model on June 13 as an open weight model. Anyone can download it, run it on ordinary hardware and change it. The release came a day after the US government blocked global access to Anthropic’s Fable 5.
Eleven days later, on June 24, China’s 360 Security Technology announced Tulongfeng, a bug detection tool it says also matches Mythos.
A challenge to export controls
The two announcements land in the middle of a month in which Washington pulled Anthropic’s strongest models offline and later allowed a limited, vetted return. It also pressed OpenAI to give its newest system only to customers the government approves. The thinking behind those steps is that limiting access buys safety and an edge.
A free Chinese model of similar skill cuts against that logic. If the capability is no longer scarce, the restrictions mostly tie the hands of American companies and do little to keep the tool from rivals.
Some caution is in order. Benchmarks measure narrow slices of ability and do not always carry over to real-world work. The tests here covered specific bug-detection tasks, not everything a frontier cyber model can do. Chinese labs also have a reason to describe their models as equal to the best American ones.
The open question for policymakers is whether export controls still make sense when the thing they target can be downloaded from a rival for nothing.
