In a new essay, AI researcher Nathan Lambert argues the discourse around cyber risks from open-weight models is broken, with policy debates trapped between two unhelpful extremes. He identifies three camps: frontier labs and national security voices who see open weights as untenable risks, Western moderates who say bans would make the world less safe, and Chinese companies that keep releasing capable open models.
Lambert takes issue with recent reporting such as Anthropic's analysis of GLM-5.3's offensive cyber capabilities, saying the technical work is reasonable but it fails to ask broader questions about what happens if open models are banned or why Chinese developers deem releases safe. He also notes that classified briefings are shaping views, while public evidence so far shows closed models have been documented in most existing AI-related cyber attacks.
He suggests the "open dangerous, closed safe" framing may be closer to "open unsafe, closed unsafe," and argues that closed-model APIs could be causing more near-term harm because of their accessibility. Lambert concludes that anyone who wants to ban open-weight models should also consider making public-facing APIs for frontier closed models illegal, since the current safeguards are far from perfect.