The debate over whether frontier AI models should be open or closed has intensified, with Anthropic's recent report arguing that the Chinese model GLM-5.3 — released in August by Z.ai and freely available for download — poses a major cyberattack risk. The report is the latest move in a long-running argument with significant commercial stakes.
Nvidia CEO Jensen Huang, who benefits directly from consumers running their own hardware to power open models, advocates enthusiastically for more open models. Anthropic CEO Dario Amodei, whose company runs a closed-model business that could be undermined by cheap, powerful alternatives, advocates for a more cautious approach. But beneath the commercial incentives, there is real substance to the disagreement.
Open models like GLM-5.3 allow researchers, developers, and governments to inspect, modify, and deploy AI without depending on a single company's API. This reduces vendor lock-in, enables customization for specific use cases, and accelerates innovation in regions that want AI sovereignty. Countries in the Gulf, for example, have invested heavily in open-model infrastructure as part of their national AI strategies.
But the risks are real. An open model can be modified to remove safety guardrails, fine-tuned for malicious purposes, or deployed in environments with no oversight. Anthropic's report on GLM-5.3 argues that the model's cyberattack capabilities are significant enough that unrestricted access creates a meaningful threat — even if the exact nature of that threat is debated.
The reality is that both sides have valid points. Open models drive competition, transparency, and accessibility. Closed models offer more control over deployment and misuse. The question is not which approach is right, but where the line should be drawn — and who gets to draw it.
As AI models become more powerful, this debate will only get more urgent. The stakes are too high for either side to claim the moral high ground.
