The US government has lifted its export restrictions on Anthropic’s new Mythos and Fable AI models, nearly three weeks after the sudden suspension sent shockwaves through the tech industry.
The news: Anthropic and US Commerce Secretary Howard Lutnick announced on Tuesday that the export controls on the Mythos and Fable models, initiated on June 12, had been lifted.
The ban was put in place after Amazon researchers reportedly found a way to bypass Fable 5’s safeguards to identify software vulnerabilities, Anthropic said. Even though Fable 5 is meant to be a restricted version of the more powerful Mythos, Anthropic claimed that the same jailbreak technique worked on other AI models, too, not just its own.
From the source: “Our testing confirmed that many less capable models … could identify the same vulnerabilities as Fable 5 did in the [Amazon] report,” Anthropic wrote in its redeployment announcement. Anthropic added that the reported jailbreak technique did not expose any “unique Mythos-level cyber capabilities” within Fable, Mythos’ sibling model that the company says has “no such unique offensive capabilities.”
Following the thread: Anthropic said despite the lack of vulnerabilities, it put in place a new guardrail for Fable 5 anyway, training a new “safety classifier” that targets and blocks use of the model that may replicate Amazon’s reported jailbreak technique. Users will now be notified if a request to Fable 5 is blocked, and their request redirected to Anthropic’s Opus 4.8 model.
Final thought: Anthropic began to re-roll out Fable to users globally on Wednesday, and said it had restored access to Mythos 5 “for a set of US organizations.”
The AI giant said it’s continuing to coordinate with the US government on expanding the number of domestic and international partners in its Glasswing program, which gives organizations access to the supposedly powerful Mythos 5. Just over a week before the US instated its export ban, Glasswing gave the Canadian government access to Mythos through the Canadian Centre for Cyber Security. BetaKit has reached out to Anthropic and the Canadian Centre for Cyber Security to clarify if Canadian organizations now once again have access to Mythos.
Feature image courtesy Anthropic via Flickr. Image license under CC BY 2.0.