Anthropic calls for threading the needle on open-source AI

Anthropic CEO Dario Amodei said the company does not support a ban on open-source AI but called for targeted frameworks to prevent illicit model distillation and development by adversarial entities, citing China as the primary threat. Amodei disagreed with claims that open-weight models necessarily improve cybersecurity, pushing back against an open letter from NVIDIA, Palantir, and Cisco that argued open-source models benefit security.

Anthropic calls for threading the needle on open-source AI Anthropic CEO Dario Amodei cited China as the primary threat to such a paradigm and cast doubt over the narrative of open-source as a cybersecurity asset. Anthropic doesn’t support a ban on open-source artificial intelligence but is wary of models that have potentially dangerous capabilities, the company said in a Monday blog post following several other tech leaders forming https://www.nextgov.com/artificial-intelligence/2026/07/over-30-companies-form-open-source-ai-alliance/415028/?oref=ng-homepage-river an open-source AI Alliance. Anthropic CEO Dario Amodei outlined his concerns in the post https://www.anthropic.com/news/position-open-weights-models about open source models, calling for “targeted legal and commercial frameworks” to prevent illicit model distillation and development by adversarial entities. “To summarize my and Anthropic’s position, we have not and are not advocating for a ban on open-weights models as a category,” Amodei wrote. “We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed.” Echoing the White House’s recent accusation https://www.nextgov.com/artificial-intelligence/2026/04/white-house-accuses-china-deliberate-industrial-scale-campaigns-steal-us-ai-models/413083/ that Chinese AI developer Moonshot AI illegally stole Anthropic's intellectual property through distillation, Amodei said his primary concern is AI developed by authoritarian governments, a risk that he said can be remedied through export controls on chips to China, mitigating large-scale distillation and safety testing on both open and closed-source models. Although he acknowledged his security concerns, Amodei noted that he understands the value of open-source models, and that they serve a public good by democratizing access to AI technology. Amodei maintained this position when addressing the open letter https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf supporting open-source model access issued last Friday by other major technology companies, including NVIDIA, Palantir and Cisco. He wrote that he agrees with many parts of the letter –– mainly that open-source models benefit the AI economy and provide user customization –– but disagreed with the letter’s stance on open-weight models' cybersecurity benefits. “I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers,” Amodei wrote. The question as to whether or not open-source models are a boon to security arose when OpenAI’s GPT-Sol 5.6 conducted an attack on Hugging Face networks https://huggingface.co/blog/security-incident-july-2026 in July after breaching containment in an internal testing environment. To identify attackers and secure their systems, Hugging Face used a popular open-source model, GLM 5.2. Open-source models’ security status is currently a pressing topic in D.C. David Sacks, the co-chair of the President’s Council of Advisers on Science and Technology, has criticized Anthropic’s reticence on open-source models, particularly surrounding training models via distillation. “Anthropic maintains that it is entitled to train for free on all the world’s output, even if the author objects,” Sacks wrote in a post on X https://x.com/DavidSacks/status/2081867338392150282 on Monday. “But if a competitor trains on Anthropic’s output after paying for it, that is IP theft. The hypocrisy is breathtaking.” NVIDIA Founder and CEO Jensen Huang also doubled down on his support for open-weight models. NVIDIA united over 30 companies https://www.nextgov.com/artificial-intelligence/2026/07/over-30-companies-form-open-source-ai-alliance/415028/?oref=ng-homepage-river to create an Open Secure AI Alliance to promote the use of open source models for cybersecurity. Following this announcement, Politico reported https://www.politico.com/live-updates/2026/07/27/congress/nvidias-huang-heads-to-capitol-hill-for-ai-talks-01012902 that both Huang and OpenAI CEO Sam Altman went to the White House to talk about AI policy writ large. “Jensen is in DC to meet with leaders on both sides of the aisle to discuss NVIDIA’s work to strengthen the country’s supply chain by producing $500B in American technology over four years and American leadership in AI, including leading in open source,” a NVIDIA spokesperson told Nextgov/FCW .