Open models have real benefits. Do they outweigh the risks?

Anthropic’s latest warning lands in a familiar place: the argument over open AI models that never really goes away. The company says GLM-5.3, a model released in August by China’s Z.ai, poses a major cyberattack risk because anyone can download it and alter it. That complaint is about more than one model. It’s really about whether the most capable systems should be open to the public at all.

The case for openness is practical as much as political. Open models can be downloaded, run on local hardware, retrained, combined, compressed, and fine-tuned. Researchers studying how AI systems work often have no real choice but to use them, because closed models only let outsiders send prompts and read the replies. Businesses like them too, especially when they need to handle sensitive or proprietary data on their own machines instead of sending it to an outside provider.

And then there’s the power question. Supporters of open models argue that a small number of frontier labs shouldn’t control the infrastructure behind a technology this important. Elizabeth Seger of the Tony Blair Institute for Global Change says the issue is as big as control over information on the internet. Hamish Hobbs of the Centre for Long-Term Resilience goes further, saying a world where individuals have their own safe AI systems looks less dangerous than one where people are told to trust an inscrutable model run by a government.

But the safety case is not imaginary. Anthropic says people have already used Claude in extensive hacking operations, and open models would be easier to abuse because they can be run without provider filters or account bans. Researchers also say it can be surprisingly easy to retrain a model that was supposed to refuse harmful requests. That makes the classic open-source argument, where many eyes catch bugs, only partly reassuring. With models, a bug can be a weapon.

So the real debate is not open versus closed in the abstract. It is how much risk society will tolerate in exchange for lower costs, more research access, and less corporate control. Open models do spread power. They also spread capability, including to people who should not have it. Both things are true, which is why this fight keeps getting louder instead of cleaner.