The Federal Register website, operated by the U.S. National Archives and Records Administration, used Alibaba’s Qwen artificial intelligence model to help users search proposed federal regulations and browse public comments. According to Reuters, the Qwen-powered search option was available on Sept. 16 and was removed the same day, around the time social media posts began drawing attention to its use. Reuters said it was unclear when the tool had first been deployed.
This incident occurred against the backdrop of recent concerns by the U.S. government and cybersecurity agencies regarding Chinese AI companies acquiring U.S. AI technology. The Federal Bureau of Investigation (FBI) had previously stated that Alibaba had used U.S. AI models such as Claude to gain technical capabilities. Anthropic also noted in its latest report that it had detected large-scale model distillation activities linked to Alibaba.
Why did a US government website use Qwen?
The Federal Register is the official publication of the U.S. government, releasing federal regulations, proposed rules, notices and other government documents from Monday to Friday.
Based on website screenshots and archived code, Reuters confirmed that Qwen appeared in the AI search option on the Federal Register website on Sept. 16. Users could use this feature to search federal regulations and public comments. It remains unclear exactly when Qwen first appeared on the website.
According to Reuters, the Qwen search feature was removed the same day after related information surfaced on social media. Neither the National Archives and Records Administration nor the White House responded to requests for comment, and the Federal Bureau of Investigation (FBI) also declined to comment.
For the best of our weekly content！
You are now signed up for our newsletter
Check your email to complete sign up
CCP AI labs distilled Claude
In September, Anthropic published a report claiming to have discovered that multiple Chinese Communist Party AI labs had accessed Claude via a large number of accounts, in an attempt to utilise its capabilities for training their own models. Anthropic defined such behavior as “unauthorized distillation” and noted that the largest operation was linked to Alibaba.
According to the report, between May and July, related activities generated over 151 million interactions with Claude using more than 3,500 fake accounts, with daily interactions peaking at almost 3 million. Anthropic believes that this data was used to help train Alibaba’s Qwen series of models.
Subsequently, the U.S. government stated that Chinese AI companies were engaged in large-scale activities to acquire U.S. AI technology. A Reuters report on Sept. 8 noted that U.S. cybersecurity and law enforcement agencies had described these activities as being on an “industrial scale.”
Did this involve US government secrets?
Currently, there is no public evidence to suggest that Qwen’s use on the Federal Register website resulted in the leakage of U.S. government secrets. This is mainly because the Federal Register website primarily handles publicly available regulations, government documents and public comments.
However, the real issue that needs to be clarified is exactly how Qwen was deployed and whether the data generated by user searches left the U.S. government’s information systems.
Experts interviewed by Reuters pointed out that the situation differs depending on whether the model runs on the government’s own systems or data is sent directly to an external AI service.
Mark Warner, vice chairman of the U.S. Senate Intelligence Committee, stated that the risk depends on whether government data left secure boundaries and was processed on systems controlled by Alibaba.
Original article
Comments