Anthropic's Amodei says open models make bioweapons too easy — and he's not backing down

Dario Amodei wants to be crystal clear: he’s never called for an outright ban on open-weight AI models. But in a new blog post, the Anthropic CEO doubles down on his conviction that highly capable open models are uniquely dangerous, framing his argument around what he calls two “nightmare scenarios.” The first is China racing ahead of the U.S. to build militarily superior AI. The second, and the one where open weights play the villain, is the potential for bad actors to grab a model with no safety guardrails and repurpose it for cyberattacks or, more chillingly, to engineer biological weapons.

Amodei’s argument hinges on an asymmetry he sees in biology. He contends that while open models can empower digital defenders, a sufficiently advanced model could dramatically accelerate the creation of a weaponized virus, whereas developing a defense—even in the best case—would take years. This isn’t a hypothetical, he says; AI labs already see misuse attempts on closed models. The difference with open weights is that once a model is released, there’s no calling it back. He points to a UK AI Security Institute study showing open-weight models trail cutting-edge systems by just a few months on cybersecurity metrics, a gap he clearly finds uncomfortably narrow.

His proposed fixes go well beyond licensing. Amodei wants stricter chip export controls to prevent China from training more powerful models without U.S. semiconductors and a crackdown on “industrial distillation,” a technique he claims lets Chinese firms act as “free riders” on American AI progress. Critics immediately push back, noting that cutting off chip sales mostly accelerates China’s domestic chip industry and that distillation is a legitimate, widely used machine learning technique. Amodei also advocates for mandatory safety testing for all powerful models, though he sees a growing consensus for exempting weaker systems from startups and researchers.

The blog post lands amid reports that U.S. officials are considering moves to restrict the use of Chinese open-weight models, a push many major tech companies recently opposed in a signed letter. The subtext here is hard to ignore: Anthropic and OpenAI, which is reportedly still lobbying against open weights despite signing that letter, are regularly accused of trying to pull up the ladder behind them to fend off cheaper competition. Amodei preempts that charge by insisting a ban on Chinese models wouldn’t stop bad actors and would only “protect US AI companies from competition,” a goal he claims isn’t his. His consistency is real—he’s warned about his own models’ capabilities before and those warnings helped shape current U.S. export bans—but the line between principled safety advocacy and commercial self-interest remains, for many, blurry at best.