US Accuses Six Chinese AI Firms of Industrial-Scale Model Distillation Attacks

On September 9, 2026, U.S. cybersecurity and intelligence agencies released a joint statement alleging that six Chinese AI companies have been conducting industrial-scale distillation attacks against American frontier AI models since at least late 2024. The attacks involve systematically querying target models to extract vast amounts of data, including training information and model capabilities, effectively 'distilling' the models' knowledge for unauthorized use.

The agencies did not name the specific companies or disclose the exact number of tokens extracted, but described the operations as 'industrial-scale' and ongoing. The accusations highlight a growing concern over the security of AI models, which are increasingly viewed as critical national security assets. Distillation attacks can undermine the competitive advantage of U.S. AI developers and potentially enable adversaries to replicate advanced capabilities without the original investment.

While the technical feasibility of model distillation is well-established, the specific attribution and scale of these alleged attacks remain unconfirmed. The agencies have not provided public evidence, and the named companies have not yet responded. The report emphasizes that the attacks are not traditional software exploits but rather sophisticated abuse of legitimate AI interfaces, making them harder to detect and mitigate.

The U.S. government is reportedly considering policy and diplomatic responses, as well as encouraging AI developers to implement stricter access controls, rate limiting, and anomaly detection to identify and block suspicious query patterns. No software patch can fully prevent distillation, but proactive monitoring and robust API governance can reduce the risk.

This incident underscores the evolving threat landscape where AI models themselves become targets. Organizations developing advanced AI systems should treat model access as a high-value asset and apply security best practices accordingly.