After Commerce Department lifts limits, Anthropic moves to plug loopholes that allegedly let some Chinese firms access Claude, reports say

Anthropic is moving to close “loopholes” that, according to a report, have allowed some Chinese technology companies to gain access to Claude, the company’s most advanced generative AI models. The move was described in reporting that cited people familiar with the matter, and it comes days after the Commerce Department lifted export controls that had been in place for roughly three weeks on certain Anthropic models.

The timing links the company’s internal access controls to a broader U.S. effort to manage national-security risks tied to advanced AI. The New York Times reported on June 30 that the Commerce Department lifted restrictions on all of Anthropic’s AI models, describing the change as the end of a short-term limitation that had been imposed over national security concerns. Under that framework, the question becomes not only whether a model is permitted to be exported, but how usage restrictions are enforced once access is technically allowed.

Prior to the June 30 shift, multiple public accounts described the government as having directed Anthropic to suspend access for certain foreign users. The New York Times reported on June 12 that Anthropic said it had been ordered by the U.S. government to suspend access for all foreign nationals to its Fable 5 model. In the months leading up to these actions, other reporting said the Pentagon cut off business with Anthropic and declared it a national security risk, which Anthropic disputed and said left it “no choice but to” comply with government direction.

The new reporting on loopholes centers on how Chinese firms may have been able to work around Anthropic’s “stringent restrictions,” as described by outlets that summarized the Financial Times account. Those reports did not identify specific companies or the exact mechanism in public versions available through search results, but they said Anthropic is taking steps to limit unauthorized use and prevent circumvention of its controls.

If implemented, changes would likely affect the compliance boundaries that Anthropic applies to third parties, including how it evaluates requests for access and how it monitors or blocks attempts to bypass stated limitations. For policymakers, the practical issue is that export and access restrictions can be undermined when intermediaries, licensing arrangements, or downstream deployment do not map cleanly onto the government’s policy goals.

Neither Anthropic nor the Commerce Department was cited in the available reporting as publishing a detailed enforcement notice for the reported tightening. The account therefore leaves open what specific changes will be made, whether they are tied directly to export-control classifications, and how quickly they will take effect for existing users.

The development is likely to renew scrutiny of how the U.S. government calibrates AI export controls, especially during periods when limits are lifted or restructured. For companies providing frontier AI, it also highlights that compliance systems may need to evolve as soon as access rules change, even when the underlying model remains the same.