US lifts export controls on Anthropic’s frontier cybersecurity AI models

Source Domain: therecord.media

Anthropic restored global access to its Fable 5 model on Wednesday, ending a roughly three-week shutdown that began with the U.S. government imposing export controls barring foreign nationals from accessing the advanced, cybersecurity-focused AI tool.

It comes as the Five Eyes intelligence alliance has warned business leaders to immediately prepare for the impact frontier AI models will have on cybersecurity, with new models set to “fundamentally transform” both offensive and defence capabilities. “The timeline is not years, it is months,” the agencies warned.

At the time the restrictions against Anthropic were imposed, the company said it had been forced to disable access for all customers to ensure compliance. In an update Tuesday, Anthropic said the controls had been lifted after the company came to a series of agreements with the government.

The episode marked the first known use of export control authorities to pull AI software rather than chips or hardware from public access. Its reversal may set the terms under which frontier AI models are regulated in the U.S. going forward.

Export controls on Anthropic’s more powerful cybersecurity model, Mythos 5, were also fully lifted as of June 30, although access to that model remains restricted to vetted U.S. organizations through Project Glasswing — Anthropic’s controlled-access program for critical infrastructure defenders. The company said it is continuing to negotiate broader domestic and international access through Glasswing.

The initial shutdown was, according to Anthropic, triggered by a “jailbreak” technique covered in an Amazon research report. That technique was subsequently described in detail by Katie Moussouris, founder of Luta Security, whom Anthropic asked to assess the paper.

Moussouris wrote that researchers fed Fable 5 open-source code with publicly known vulnerabilities plus deliberately planted flaws, then asked it to “fix this code.” The model’s output was then…

Source