As expected, Anthropic has finally issued an official stance on open-weight models.
Against this backdrop, there has been a broad rallying of the American tech sector in favor of these models, in a context that is pushing Washington to curb access. Nearly a hundred organizations — including Google, Meta, Microsoft, NVIDIA and OpenAI* — signed an open letter praising the benefits of open-weight for the economy and society. Anthropic is not part of the group. Some observers saw this as a form of support for the administration’s ambitions, aimed at protecting a business model that relies exclusively on selling closed models.
The open-weight—a bearer of irreversible risks?
Dario Amodei rejects this reading of events. Banning open-weight models isn’t the solution, says the head of Anthropic. He even says he is in favor of it… provided they do not pose a danger. He cites his essay “The Adolescence of Technology,” written earlier this year, where he expresses concern that “authoritarian governments” could develop AIs “more powerful than those of the United States” and use them to “enhance their military, intelligence and surveillance capabilities.” These concerns align with those voiced by J.D. Vance at the AI Summit 2025.
Dario Amodei also fears the misuse of AI for cyber or biological attacks. On this point, open-weight models pose a greater risk than closed models, he argues. The reason: it is very hard to apply guardrails or monitor their use, and once the weights are published, they cannot be recalled. To back his claim, the executive cites a comment from the British AI Safety Institute, which notes that opening weights “creates a persistent and irreversible risk of misuse” by making it possible to “definitely lose control over safeguards and user access.”
Anthropic calls (again) to curb access to chips and to distillation
A simple ban on the use of open-weight models would not solve the problem, says Dario Amodei. It would indeed protect companies like Anthropic, he concedes. “But that has never been my objective …”
For him, the main remedy would be not to sell China powerful chips, nor equipment to manufacture those chips. All while acting in parallel against circumventing export restrictions. The head of Anthropic has long held this view. He believes China’s production capacity is limited: scale laws imply that without American chips, it cannot develop models more powerful than the United States.
The second lever, also long proposed, is to push politically against “distillation industrial operations.” Amodei adds mandatory safety tests for all models “sufficiently capable,” whether open or closed, regardless of their country of origin. That implies Beijing would need to cooperate. Anthropic considers that limited cooperation is possible on the risk of designing biological weapons, since China has an interest in this area.
Risks of attacker–defender asymmetry
Dario Amodei agrees with the open-letter signatories on the idea that the answer to the distillation problem should come from targeted legal and commercial frameworks. He is broadly in agreement on the economic and societal benefits of AI, and, “at least in some cases,” on giving users greater control.
Yet the executive does not think openness necessarily fosters the development of guardrails. Nor that in cyber affairs it benefits defenders more than attackers. “The opposite is at least as likely,” he argues, illustrating with a potential strong asymmetry in biology. Sufficiently powerful models could quickly turn viruses into pandemic-level weapons, while defending against such viruses would be a multi-year effort.
* OpenAI did not sign the letter immediately.
