The U.S. State Department escalated its response to intellectual property theft in artificial intelligence last week, issuing a diplomatic cable that for the first time names specific Chinese firms—DeepSeek, Moonshot AI, and MiniMax—as perpetrators of large-scale model extraction and distillation. The April 24 directive instructs U.S. envoys to brief foreign governments on what the department describes as adversarial campaigns to steal American AI models through query-based distillation attacks.
Distillation works by sending millions of queries to a target model via fake accounts and proxies, recording responses, then training a smaller, cheaper model on those outputs. The technique bypasses the need for massive datasets or expensive GPU clusters. Anthropic documented the scale in February: DeepSeek, Moonshot, and MiniMax operated 24,000 fraudulent accounts, generating over 16 million exchanges with Claude. MiniMax alone accounted for 13 million queries. Moonshot targeted agentic reasoning, coding, and vision tasks with 3.4 million hits.
OpenAI reported similar patterns to lawmakers, describing DeepSeek’s “free-riding” on their systems. White House OSTP Director Michael Kratsios followed with a memo to federal agencies, stating that foreign entities—primarily in China—are running “deliberate, industrial-scale campaigns to distill U.S. frontier AI systems.” Proxy accounts evade rate limits; jailbreaks extract proprietary behavior. The result: Chinese models like DeepSeek-V3 and R1 match GPT performance at a fraction of the cost, sidestepping U.S. chip export controls entirely.
The cable warns allies against using models distilled from U.S. proprietary systems and lays groundwork for potential sanctions. Lawmakers are exploring violations under the Economic Espionage Act and Computer Fraud and Abuse Act. Bipartisan bills targeting the accused firms are advancing, and entity list additions appear imminent.
China’s embassy in Washington pushed back, stating Beijing “attaches great importance to the protection of intellectual property rights.” None of the named companies commented. The timing is consequential: President Trump is scheduled to meet Xi Jinping in May, and this cable hardens the U.S. negotiating position.
Industry response has been mixed. U.S. labs collect API fees during attacks—tens of millions—but lose competitive advantage as Chinese models flood markets at low prices. DeepSeek’s V4 preview, launched Friday amid the controversy, runs on Huawei Ascend chips, signaling self-reliance. Yet insiders note these distilled models falter on complex reasoning, hallucinate more, and lack generalization depth. Benchmarks deceive; real-world performance lags.
For engineering teams, the takeaway is architectural. Rate limiting, anomaly detection, and geo-fencing remain first-line defenses, but at 16 million queries, scale defeats most static rules. Model-level watermarking and output fingerprinting are emerging as necessary countermeasures. The U.S. government is now coordinating intelligence sharing with labs to identify extraction patterns and harden systems before the next wave.
This isn’t a legal gray zone anymore—it’s a full diplomatic confrontation over AI sovereignty. The cable makes clear: distilled models are not just cheaper copies; they are stripped of safety rails, posing risks in cybersecurity and biosecurity. Enterprises evaluating Chinese models should verify provenance and understand that benchmark scores do not capture the absence of alignment training.
As the May summit approaches, the AI cold war has a new front. Models are now munitions, and distillation is the weapon of choice.
Source: Webpronews