Open AI Models: Democratization vs. Cyber Threats

The Open Model Debate Heats Up

Anthropic released a report labeling GLM-5.3 (a model from the Chinese company Z.ai, released in August) a serious cybersecurity threat. The core concern: the model is freely distributed, and anyone can download and modify it. This is the latest move in a long-running debate over whether powerful AI models should be open or kept under the control of major labs.

Behind the argument sits a significant commercial stake. Nvidia CEO Jensen Huang champions open models — his company benefits when people buy their own hardware to run them. Anthropic's Dario Amodei takes a more cautious stance, since cheap open models threaten Anthropic's business model. But commercial interests mask real disagreements about how the future of AI should be shaped.

Two Sides of the Argument

David Gray Widder (University of Texas at Austin) notes that which side you land on depends on your overall view of AI safety. Safety advocates fear that open models are too easy to misuse. Access advocates insist that such a powerful technology shouldn't be monopolized by a handful of companies.

An open model (like GLM-5.3) consists of files you download and run on your own hardware. A closed model (ChatGPT, Claude) is only accessible through an API or web interface. Today, closed models are more powerful — the top 8 models on the Artificial Analysis LLM leaderboard are all closed. But open models trail by only a few months of development.

Advantages of open models. Researchers can see a model's internal workings and test hypotheses about how it operates. Businesses save money: open models are cheaper than closed ones, according to a Mozilla report. Local execution protects sensitive data — companies can process confidential information on their own hardware without handing it to third parties. Elizabeth Seger (Tony Blair Institute for Global Change) sees the main upside as weakening major labs' monopoly on one of the era's defining technologies. Hamish Hobbs (Centre for Long-Term Resilience) adds that a world where everyone has access to powerful, safe AI is far less dangerous than one where a single government controls all the most advanced hacking and cyberattack models.

But Paul Azunre (Ghana NLP) warns of a downside: large companies take improvements from open projects and fold them into their own products without crediting the developers. Mark Zuckerberg has openly said that the ability to gather ideas from the community and integrate them into Meta is a key benefit of open models.

Risks of open models. Anthropic published a report on hacks carried out using Claude. In July, OpenAI's agents were used to breach Hugging Face. Closed models screen incoming requests for danger (hacking, bioweapons, and the like) and can suspend a user's account. But once you've downloaded a model to your own machine, there's no filter and no way to revoke access. Researchers have shown that it's easy to retrain an open model to agree to things it previously refused. If a very powerful open hacking model were released, an attacker could cripple critical infrastructure before people had time to defend against it.

But the situation is more complicated than it seems. After the breach, Hugging Face analyzed it using an open model, because closed models blocked questions about the hack itself. Some experts draw a parallel to Linux and open-source software: they're considered safer thanks to transparency and thousands of independent developers hunting for vulnerabilities. Open models can be used for cyberdefense too — the same tool that enables an attack can also enable protection. Anthropic makes its models available for cyberdefense through Project Glasswing, but only to select companies. Open models are more universally accessible in this regard.

Still, most experts view open models as the riskier option. Elizabeth Seger notes that the main danger lies in rare but catastrophic events. If a model with its own goals starts causing harm, decentralization will make the problem harder to detect.

Searching for a Middle Ground

For a long time, the debate was binary: either full openness or full safety. But there's growing recognition that options exist between an outright ban on open models and recklessly releasing them. Any approach requires a trade-off between protection and democratization. Making sound decisions about whether to support or restrict open model development requires understanding these risks and benefits in their full depth.