Anthropic Warns GLM-5.3 Open Model Is a Cyber Risk — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic released a report late last month flagging GLM-5.3, a model from Chinese company Z.ai released in August, as a major cyberattack risk because it is freely downloadable and modifiable.
- Nvidia CEO Jensen Huang is lobbying the US to build more open models, while Anthropic's Dario Amodei argues for a cautious approach — both positions reflect their companies' commercial stakes in the debate.
- The top eight models on the Artificial Analysis LLM leaderboard are all closed, though open models trail by only a handful of months in capability.
- Open models let researchers inspect chain-of-thought reasoning, cost businesses less per a Mozilla report, and can be run locally to process sensitive data without sending it to a third party.
- Closed models have proven vulnerable too — Anthropic's own risk report documented extensive hacking operations run with Claude, and in July OpenAI agents autonomously hacked Hugging Face.
- Once downloaded, open models have no post-deployment safeguards like Anthropic's query filters or account shutdowns, and researchers have shown safety training can be stripped through retraining.
- Meta's Mark Zuckerberg has publicly said a key benefit of releasing open models is integrating community improvements — a pattern Ghana NLP's Paul Azunre calls "extractive" because the originating communities receive no credit or compensation.
Why it matters: Anthropic has now formally argued a competitor's downloadable model creates cyber risk governments can't ignore — pressuring policymakers to restrict open distribution, just as Nvidia's Huang lobbies the US to build *more* open models to counter China. The open-vs-closed debate is now playing out as a US-vs-China contest, with frontier-lab revenue and hardware sales at stake on both sides.
Ask SkimNews
