Z.ai GLM-5.2 Matches US AI on Cybersecurity Bugs
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Z.ai's GLM-5.2 matches latest US models in finding security bugs, according to researchers cited by the Wall Street Journal, intensifying scrutiny of US export policy toward Chinese open models
- The Wall Street Journal frames the finding as evidence that Washington's clampdown on top US AI is fueling concern the US is handing Beijing a cyberwarfare advantage
- A Jefferies report (covered by Business Today and China Money Network) says China's cheaper AI models are quietly taking over enterprise AI, putting price at the center of the model race
- The Information reports the cybersecurity match specifically against Anthropic's "Mythos" model
- David Sacks, posting from inside the administration on X, defends Trump's AI strategy as pro-innovation, pro-infrastructure, pro-energy, and pro-export, warning that deviating from it is at the US's peril
- X user Tim Hua publicly disputes the WSJ framing, offering to bet $100-to-$1 that GLM-5.2 will score below Mythos and GPT-5.5 at UK AISI's cyber range
- X user Kyle Russell reports internal benchmarking where GLM 5.2 was judged "a bit better" than Opus 4.8 at less than half the cost for updating a mortgage servicing knowledge base
Why it matters: The benchmark equivalence claim is feeding directly into the US export-control debate, with Sacks citing administration strategy from inside the government while the Jefferies report frames Chinese open models as an enterprise threat. The technical dispute — Hua willing to bet $100-to-$1 the equivalence doesn't hold on independent UK AISI testing — is the wrinkle the policy narrative papers over: if GLM-5.2 doesn't actually match Mythos on adversarial testing, the cyberwarfare-advantage argument collapses.
