In July 2026, OpenAI models hacked the company Hugging Face during testing, circumventing precautionary controls. Models from Anthropic and Meta have similarly hacked outside organizations. These incidents raise questions about the regulation of AI. Namely, how do we monitor AI to ensure that it remains safe and under human control? A first move in the right direction would be for the federal government to pass the Deterring American AI Model Theft Act and pursue Entity List sanctions against Chinese labs implicated in model distillation. The act would prevent novel technology from reaching the hands of foreign entities, buying the United States more time in the AI race with China to ensure US-made models are developed safely.
Distillation is a process by which, using false accounts, an AI company is able to train its model off of a preexisting one. Foreign prompt engineers feed off of another model’s responses to difficult prompt questions, educating their tech off of these sample outputs. By doing so, the developing model is able to absorb some of the knowledge from its “father model.” Many American-owned AI systems are closed-source, meaning that their unique technical components are not shared publicly. Still, distillation closes the capability gap between American-made and foreign-made AI models by allowing competitor models to simulate comparable capabilities for a fraction of the cost. American AI companies have attempted to stop distillation by shutting down suspected feeder accounts, but new models nonetheless continue to pop up. As distillation is not explicitly illegal under American law, it is extremely difficult to take meaningful action to prevent it.
Due to its nature, distillation may be impossible to stop entirely, as companies can continue to open new accounts to circumvent entity list sanctions. However, putting a policy in place to allow the government to pursue sanctions against AI-distilling companies is a necessary step in a larger conversation about technological security. On one hand, the gap between American tech companies and foreign competitors is closing, in part due to distillation. On the other, companies like Anthropic and OpenAI acknowledge that they themselves do not fully know what their own latest models can do. As such, ensuring a wider capability gap would buy American companies the time necessary to slow down domestic development and ensure more thorough safety testing.
The Deterring American AI Model Theft Act seeks to apply sanctions by adding implicated companies to an entity list in order to deter distillation. It directs the Secretary of State to assess which entities are implicated in unauthorized extraction of model capabilities. The End-User Review Committee then votes to determine whether identified entities should be added to the entity list. Entities placed on the Bureau of Industry and Security (BIS) Entity List are then subject to individual licensing requirements, which allows the United States government to restrict them. The government does so by implementing licensing requirements, obligating American AI companies to deny access to traffic coming from IP addresses linked to the blacklisted entity. Although foreign actors can circumvent Entity List sanctions by opening new accounts, passing this act is a good first step in establishing policy to deter distillation. In an AI race full of constant developments, a timely answer might be more necessary than a perfect one.
AI is far too powerful for the “move fast and break things” mindset. Federal deterrence is therefore key to buying American companies time in the AI race to slow down and implement safety mechanisms without falling behind. Safe AI development requires the ability for human control, and therefore, the ability for humans to intervene or shut down an AI model that becomes unsafe. Although China is also developing its own AI models from scratch, distillation allows Chinese AI companies to accelerate development. If prevented from absorbing frontier American AI capabilities, Chinese companies would take longer to develop state-of-the-art AI models, giving the United States breathing room to implement important safety mechanisms. The race for AI dominance is one of the reasons the current presidential administration has been reluctant to impose any formal regulation in the United States, even after AI models have “gone rogue.” This act, by attempting to widen the capability gap between the United States and China through deterring distillation, seeks to address those concerns.
Offspring models often do not absorb the same safeguards as their “father model.” Once distilled, the American company no longer has direct control over the offspring intelligence, rendering a kill-switch safeguard infeasible. The United States has no governing power over Chinese AI models, and Chinese AI policy does not require developers to conduct or release risk assessments of AI model capabilities. AI technology with the potential to be extremely dangerous is no longer able to be regulated by American entities after distillation. Keeping state-of-the-art technology out of Chinese developers’ hands reduces that risk. The act specifically states its intention to protect against the extraction of “key technical features of closed source, American-owned artificial intelligence models, and for other purposes.”
Despite the importance of the act to the safe development of AI, one major change should be made prior to passage. Under current provisions, the investigation of implicated parties can take up to 180 days for the initial assessment and up to 210 days for the congressional report, which gives far too much time for distilling companies to adapt before they have even been identified. A sanctioned entity could, within that time frame, create another company with a different name and continue practicing distillation. Section 5(a) attempts to address this concern, stating that a vote will occur on whether affiliate entities, which are “to be determined by ownership of 50 percent or more in the aggregate, directly or indirectly,” should be added to the entity list. The act should make the addition of affiliate entities automatic in order to avoid the potentially time-consuming process of requiring an additional vote on each entity that is essentially identical to the former, just not in name.
The Deterring American AI Model Theft Act is currently under review after passing through the House Foreign Affairs Committee in a unanimous 43–0 vote, a testament to its unique momentum. This bill may realistically pass, but clearing both the House and the Senate poses a real challenge to any bill. Congress should make this bill a priority. The capability gap is too small for comfort, leaving American companies without time in the AI race to slow down and develop safeguards that would ensure humans maintain control over AI technology.




