The US AI Safety Regime Keeps Producing Its Own Opposite

The Pentagon designated Anthropic a national security supply-chain risk in late February, a label normally reserved for foreign adversaries. The designation followed CEO Dario Amodei's refusal to remove the guardrails that block Claude from autonomous weapons and mass surveillance. The same military that blacklisted the company continued using Claude for intelligence and targeting during operations in Iran and the capture of Nicolás Maduro in Venezuela. ^[1] The White House ordered reviews of advanced AI models after Anthropic's Mythos model penetrated classified government systems in a testing exercise. ^[2] The same administration rejected global AI governance at the New Delhi summit in February. The US position, delivered by representative Michael Kratsios, was that the country rejected any binding international framework for AI. ^[3] The result is a sovereignty model in which each nation regulates itself: the US restricts its own models through review orders while Chinese models face no binding international framework. White House AI adviser David Sacks called restrictions on Chinese open-source models a regulatory capture strategy designed to protect a duopoly of OpenAI and Anthropic. He warned that using regulatory uncertainty as a competitive weapon should be unacceptable. ^[4] The same week, the Treasury Department threatened sanctions against Chinese AI firms for IP theft via distillation. ^[5] When Hugging Face, the New York-based AI startup, needed to analyze a security breach, it turned to its own AI models first. OpenAI and Anthropic models refused the task: their safety guardrails blocked hacking-related work. The company then used Zhipu AI's open-source GLM-5.2, which completed the analysis without restriction. OpenAI scrambled to add Hugging Face to its trusted access program afterward, an implicit admission that the default guardrails had blocked legitimate defensive cybersecurity work. ^[6] The Hugging Face case is not an outlier. Approximately 80% of US AI startups now use Chinese open-source base models, according to the US-China Economic and Security Review Commission. ^[7] Pinterest uses DeepSeek R-1 for its recommendation engine, reporting 30% better accuracy than proprietary alternatives. Airbnb runs Alibaba's Qwen for customer service agents. ^[8] The vacuum cannot close toward permission. US firms keep gating their most capable tools behind permissioned access programs. Anthropic withheld its Mythos model from public release entirely, creating Project Glasswing for controlled access. OpenAI launched a tiered Trusted Access for Cyber program. Google released Gemini 3.5 Flash Cyber exclusively to governments and trusted partners. ^[9][10] The two leading US AI firms publicly attack each other's safety postures: Anthropic's CEO called OpenAI's Pentagon safeguards safety theater and straight up lies, while Sam Altman called Anthropic's critiques doublespeak. ^[11] Senior safety researchers resigned from both companies in February. ^[12] The vacuum cannot close toward restriction either. Chinese models face none of these constraints. They have outpaced US models in global token usage for five consecutive weeks, processing 12.96 trillion tokens versus 3.03 trillion for US models. ^[13] DeepSeek permanently cut prices by 75%, making its models 12 to 19 times cheaper than equivalent OpenAI and Anthropic offerings. ^[14] GLM-5.2 was trained entirely on Huawei Ascend 910B processors and costs less than one-tenth of Anthropic's Fable 5. ^[15] NIST tests found that DeepSeek-based AI agents were 12 times more likely to follow malicious instructions than US models. ^[7] The safety gap between US and Chinese models is measurable and real. It is also, for the 80% of US startups and the major enterprises that have adopted Chinese open-source models, the point.