U.S. Government Lifts Its Block on Anthropic's AI Models — With Strings Attached

Key points

• The U.S. Commerce Department banned foreign nationals from using two Anthropic AI models on June 12, forcing the company to pull both products for all users worldwide.
• Amazon researchers found a method to bypass safety filters in Anthropic's Claude Fable 5 model that could allow the AI to find and exploit software flaws in real systems.
• As of Tuesday, Claude Fable 5 is publicly available again; the more powerful Mythos 5 remains limited to U.S. Organisations approved by the federal government.
• Rival AI company OpenAI restricted its new GPT-5.6 Sol model to government-approved customers following a similar request from the Trump administration.
• President Trump signed an executive order last month allowing the government up to 30 days to review advanced AI systems for national security risks before public release.

Anthropic, the San Francisco company behind the Claude family of AI chatbots, spent several weeks shut out of its own product launch. Now it's getting most of that access back, though not all of it, not without conditions.

The trouble started June 12, when the U.S. Commerce Department blocked foreign nationals from using two of Anthropic's newest AI models. That order forced the company to pull both products entirely, for every user everywhere, just days after launch.

The reason: researchers at Amazon, which provides the computing infrastructure Anthropic runs on, had found a method of bypassing Fable 5's built-in safety guardrails, the rules designed to stop the AI from doing harmful things. With those guardrails gone, the model could locate weaknesses in software systems and potentially help someone exploit them.

How dangerous is an AI that can find software flaws?

Very, if it falls into the wrong hands. Finding holes in software, what security professionals call vulnerabilities, normally requires significant expertise and time. An AI that does it quickly hands that capability to anyone, including criminals or foreign intelligence services.

Anthropic had already flagged this concern earlier this year, warning that Mythos 5 was unusually capable at spotting those flaws. As we reported on 16 June, a coalition of security executives pushed back hard on the export restrictions that followed, arguing the block handed adversaries a strategic gift. Amazon's findings appear to have pushed the government from watching closely to acting fast.

The restrictions tracked with an executive order President Trump signed last month. That order created a framework letting the federal government review the most advanced AI systems for national security risks for up to 30 days before public release. Participation is described as voluntary, though the framework itself has not yet been fully developed.

OpenAI, maker of ChatGPT, also complied. The company said Friday its new model, GPT-5.6 Sol, will stay limited to government-approved customers for now.

For ordinary people who use Claude for writing or coding help, the practical upshot is simple: the standard version is back. The advanced Mythos 5 model remains restricted, with no public timeline for when that changes.

The bigger question isn't access. It's whether voluntary review frameworks with hard deadlines are the right tool for models that Anthropic itself has described as capable of threatening critical networks. That gap between a 30-day clock and an open-ended safety problem is worth watching.

If your organisation was evaluating either model, check with your vendor contact about current access tiers and think carefully about what data you feed any AI system while these safety questions remain unsettled.