Federal Register searched comments with Alibaba’s Qwen

The Federal Register, the National Archives site that publishes a daily list of newly proposed federal regulations, offered a search of public comments powered by Alibaba’s Qwen model. The tool was still live on September 16, a week after the FBI accused Alibaba of “malicious” copying of Anthropic’s models.

Reuters reported that the Qwen option sat alongside the site’s other search options and came down that Wednesday, around the time social media posts about it appeared. Reuters based the timeline on screenshots and an archived copy of the site’s source code. When the search first went live is unclear.

The FBI made its accusation the week before, and US officials said at the time that Chinese AI companies were copying US technology on an “industrial scale.” The accusation raised tensions between Washington and Beijing ahead of a meeting between the two countries’ leaders scheduled for the week after the takedown. A spokesperson for the Chinese embassy in Washington called the copying allegations “unfounded.”

The copying in question is distillation: a lab queries a stronger model, collects its answers and trains its own model to reproduce them. Anthropic has already made a similar charge against Moonshot AI’s Kimi K3. With Alibaba, the accusation now comes from the FBI, and the accused model turned up running on a federal website.

Daniel Castro, president of the Information Technology and Innovation Foundation, said he saw a disconnect between the US-China race to build better models and a US agency picking a Chinese-made one. “It’s one of the most insane things I’ve ever seen,” he said.

Several of the experts Reuters spoke to said the deployment may not have posed a national security risk, and the reason is the kind of model Qwen is. Its weights are public, so developers can download Qwen, adapt it for a specific job and run it, often for less than closed models from Anthropic or OpenAI cost. Security experts point out that an open-weight model can run on an organization’s own infrastructure, so sensitive data stays in-house and never goes to an outside provider. And this tool worked only on public comments, which anyone can already read on the site.

Georgetown Law professor Anupam Chander said the use of Qwen did not appear to present an immediate cybersecurity risk, though he added that this depends on how the model was trained. Senator Mark Warner of Virginia, the Democratic vice chairman of the Senate Intelligence Committee, put the question in terms of where the model ran. In a statement, he said the risk depends on whether US data left the government’s security boundary and was processed on systems Alibaba controls. The archived code Reuters reviewed showed that the Qwen search existed. Reuters did not report where the model was hosted.

The Federal Register is not the first US organization lawmakers have questioned over Qwen. In April, the House committees on China and on homeland security wrote to Airbnb, which also uses Qwen, asking how it uses the model. The letter was part of an investigation into the “national security risks created by the integration” of Chinese AI tools into platforms Americans use.

The House China Committee’s chairman, Michigan Republican John Moolenaar, wants no exceptions for government sites. “No federal government entity should use a Chinese AI model,” he said. “Doing so only makes the federal government more dependent on Chinese AI models and that is not in the national interest.”