Anthropic's Claude models cleared by U.S. after weekslong cybersecurity ban, but federal AI vetting regime continues

The United States has cleared Anthropic's newest Claude models for general release, ending a weekslong Commerce Department ban triggered by a cybersecurity finding from Amazon's cloud division. One of the models is now widely available again, while a second, more capable Claude model is being restored only to U.S.-based organizations approved by the federal government. OpenAI's parallel restriction on its newest GPT release shows the machinery behind the Anthropic block is now an operating part of U.S. AI policy.

The chain of events began on June 12, when the Commerce Department moved to block foreign nationals from using two newly unveiled Anthropic models, forcing the company to pull the products days after launch, according to ClickOrlando and the AP. The trigger was a report from cybersecurity researchers at Amazon Web Services, Anthropic's primary cloud provider, who described a method of bypassing safeguards in one Claude model that could discover and potentially exploit software vulnerabilities, the BBC reported.

Anthropic had earlier warned that one of its models was unusually adept at finding software flaws in ways that could be weaponized against critical computer networks, the company said in a blog post about the model. That combination, an external research finding paired with Anthropic's own description of capability, appears to have been enough for the Commerce Department to act before the products reached a global audience.

The lift, first reported widely on Tuesday, came after Anthropic worked with the federal government to address the concerns, according to Fox Business. The widely available Claude model is now back in general release, Anthropic said in a separate post. The more capable Claude model is being restored to a select group of U.S.-based organizations approved by the federal government, Wired reported.

OpenAI followed with its own move. The company said Friday it is restricting release of its newest GPT-style model at the request of the Trump administration, limiting access to government-approved customers for a temporary period, ClickOrlando reported.

What links the two actions is the executive order President Trump signed roughly a month earlier. That order established a framework under which the federal government can vet the national-security risks of the most advanced AI systems for up to 30 days before public release, NextGov reported. Participation is described as voluntary, but neither OpenAI nor Anthropic is releasing its frontier model without federal signoff, trade publication Export Compliance Daily reported.

The framework itself remains unfinished. The 30-day review has no formal rule book yet, CIO.com reported, and the line between voluntary cooperation and effective mandate is being drawn in real time by which companies feel they can refuse.

What is still missing is the rule that makes the precedent stick. Whether the federal government now has, in practice, a gate on which AI models reach the global market, or whether each frontier release will continue to be negotiated case by case, the next test will be whichever model next hits the 30-day clock. The Anthropic and OpenAI episodes suggest the gate is already there, built from a single cybersecurity report, a single executive order, and two of the most powerful AI labs in the world choosing to wait.