Anthropic clamps down on Chinese access as US lifts AI export bans

After a three‑week shutdown, the US Commerce Department lifted export controls on Anthropic’s frontier AI models, restoring global access to the cybersecurity‑focused Fable 5 model and ending restrictions on the more powerful Mythos 5, which remains available only to vetted US organisations through the company’s Project Glasswing programme.

The controls had been imposed after a “jailbreak” technique demonstrated that the models could autonomously find, fix, and test software vulnerabilities, marking the first known use of export authorities to block AI software rather than hardware. Anthropic responded by adding a safety classifier that blocks the technique, opening a bug‑bounty program, and agreeing to share pre‑release models with US officials.

Concurrently, Anthropic is tightening enforcement to prevent Chinese firms from accessing its Claude models. A report says companies such as Ant Financial, Alibaba‑linked entities and ByteDance used overseas subsidiaries, cloud services and “transfer stations” to bypass the ban. Anthropic is targeting these work‑arounds and updating its terms‑of‑service enforcement to block access from unsupported regions, including mainland China.