Open models have real benefits. Do they outweigh the risks?
Anthropic's report on GLM-5.3 lands at an awkward moment for anyone who believes open weights are the future. The model in question, from a Chinese lab, is genuinely useful for researchers, and the whole open-source ecosystem runs on the assumption that sharing weights is how we make AI safer and more democratic. But the cyberattack scenario Anthropic sketches is specific enough to make you pause: a capable model that can be fine-tuned without oversight, used to find vulnerabilities or craft phishing campaigns at scale.
I think the honest position is that both sides are right, and that's the problem. Open models have delivered real benefits, from reproducibility to local control, and closing them off would concentrate power in a handful of labs. Yet the genie problem is real: once weights are public, you cannot recall them. The debate tends to get framed as freedom versus safety, but the more useful question is what kind of harm is actually plausible and what mitigations, like monitoring or usage limits, could survive contact with a determined adversary.
I'd want to see more evidence before treating any single model as a decisive threat. But I also suspect the window for answering this question with evidence rather than fear is closing fast.