Regarding rumors that U.S. government agencies are considering banning domestic companies from using open-source AI models from specific countries, Anthropic co-founder and CEO Dario Amodei publicly issued a statement, clearly drawing a line: Anthropic has never proposed or advocated for a ban on open-source models.

Recently, discussions about banning open-source models have sparked intense reactions in the industry, with multiple tech companies signing an open letter supporting the open-source model. Some even questioned whether Anthropic was trying to use administrative bans to suppress competitors and protect its own commercial interests. Amodei clarified that open-source models, as long as they do not possess dangerous capabilities, are a highly valuable public good—low-cost and capable of significantly driving enterprises, developers, and researchers. Protectionist bans neither address real national security concerns nor prevent the risk of industry monopolization, which is certainly not Anthropic's position or goal.

In the article, Amodei broke down his genuine concerns about AI technology safety and national security into two layers. The first layer is the risk of authoritarian regimes using more powerful AI systems to gain permanent military advantages or deepen internal surveillance. He pointed out that in the face of such risks, whether a model is open source or whether U.S. companies use these models are not key factors—the most dangerous ones are secret models developed in secret and only used by the military or intelligence departments. The second layer is the risk of powerful AI models being maliciously misused for large-scale cyberattacks or the development of biological weapons. Although open-source models, due to their lack of continuous security protection and monitoring, pose higher challenges in security governance once released, simply banning domestic companies from using compliant open-source models does nothing to stop malicious actors, and therefore offers no real help in reducing such threats.

To truly address these risks, Anthropic has proposed three practical policy recommendations and regulatory initiatives. First, strictly limit critical computing power from flowing to potential threat actors—continue to impose export controls on high-performance chips and manufacturing equipment, and crack down on all forms of illegal smuggling and subcontracting, cutting off the computing power foundation for high-risk model training at the source. Second, combat large-scale "model distillation" (Distill) across borders. Some organizations are using distillation and other low-cost methods to extract the capabilities of cutting-edge closed-source models, significantly shortening R&D cycles; against this practice of exploiting loopholes to bypass computing power restrictions, stronger legal and commercial penalties should be established. Finally, implement comprehensive and mandatory pre-release safety testing. All models with high capability levels, regardless of whether they are open-source or closed-source, must undergo strict testing and evaluation covering cybersecurity, biosecurity, and alignment risks before being delivered to the public or specific institutions—model security should be based on objective and rigorous empirical testing rather than subjective assumptions.

Amodei emphasized in conclusion that the industry should not get stuck in the endless internal struggle between "open source and closed source," but should refocus on computing power control, regulation of unauthorized distillation, and the establishment of global unified safety testing standards, so that artificial intelligence can continue to advance under controllable and safe conditions.