Our position on open-weights models
Anthropic News · 2026-07-27
Anthropic CEO Dario Amodei clarifies that Anthropic has never advocated for banning open-weights AI models, while laying out the company's actual national security positions: chip export controls, crackdowns on industrial-scale distillation, and mandatory safety testing for all sufficiently capable models.
Appears in
Extraction
Topics: open-weights-modelsai-policyai-national-securityai-safety-testingchip-export-controls
Claims
- Anthropic has never advocated for a ban on open-weights models and does not consider such bans an effective national security measure.
- Amodei's primary security concern is authoritarian governments building AI systems more powerful than US equivalents for military superiority or domestic repression, a risk unrelated to whether model weights are open.
- Industrial-scale distillation of frontier US models by adversary states is more strategically dangerous than open-weight releases because it enables capability gains without proportional compute investment.
- All sufficiently capable AI models—open or closed, regardless of country of origin—should undergo mandatory safety testing before release to assess cyber, biological, and alignment risks.
- Biology poses a strong attacker-defender asymmetry in which sufficiently capable AI may enable rapid weaponization of pandemic-level pathogens, making proactive safety testing more critical than open-weights bans.
Key quotes
Open-weights models that don't have dangerous capabilities are a public good: they don't cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.
My primary concern is the risk that authoritarian governments—not solely the Chinese Communist Party (CCP), although the CCP is clearly the most capable threat—build AI models that are more powerful than those built by the US, and use them to achieve permanent military superiority or perpetrate incredibly deep repression of their own people.
I worry that biology will have a strong attacker-defender asymmetry, where sufficiently capable models may be able to quickly weaponize pandemic-level viruses with widely available materials, whereas defense against these agents is a multi-year operational task in the best case.