A new open-weight model from the Chinese company Z.ai, GLM-5.2, has become the focus of warnings from Western security officials and researchers, who say its coding and vulnerability-finding abilities approach those of the most capable commercial systems while being freely downloadable by anyone.
The concern marks a shift in the AI debate from capability to national security. Open-weight releases have long been credited with accelerating research and lowering costs. Officials now argue that the same openness can hand advanced offensive capabilities to state-backed hackers and criminal groups.
What GLM-5.2 can do
Z.ai, formerly known as Zhipu AI, released GLM-5.2 under a permissive license that allows anyone to download, modify and deploy it. According to Axios and eWeek, the model’s performance on software engineering and security tasks, including finding bugs in code, is close to that of frontier models such as Anthropic’s Mythos line, whose cyber capabilities prompted Anthropic to restrict access.
That comparison is central to the alarm. Anthropic’s most capable models are offered through controlled services with safety classifiers that screen requests. An open-weight model carries no such guardrails once it is downloaded, and any safety training it ships with can be removed or fine-tuned away by a determined user.
The warnings
Security officials in the United States and allied countries have said that powerful open models could make cyberattacks more effective by automating vulnerability discovery and exploit development, tasks that previously required scarce expertise. Axios reported on June 25 that the release had accelerated worries among government and industry security specialists about AI-assisted hacking.
Researchers cited in that coverage said discussions had already appeared on Russian-language underground forums about stripping GLM-5.2’s safety limits to extract more capable offensive tooling. That does not mean the protections have been defeated at scale, but analysts view the chatter as an early indicator of where criminal interest is heading.
The model’s availability through third-party hosting providers has also widened access for organizations that lack the hardware to run a very large model themselves.
An old debate, sharper stakes
The episode has reopened a question that has divided the AI field for years: whether openly released models are essential to innovation and independent research, or whether, past a certain capability level, they amount to distributing dual-use tools without controls. US export restrictions on advanced chips were designed in part to slow China’s progress toward frontier models; GLM-5.2’s performance suggests those controls have not prevented Chinese labs from reaching near-frontier capability and sharing it openly.
Policymakers in several countries are now weighing whether open-weight models above a capability threshold should face disclosure requirements, pre-release evaluation or other safeguards. How that debate resolves is likely to shape AI policy well beyond this one release.
