Anthropic said the incident revealed that the AI industry lacks a consistent approach to evaluating and responding to jailbreaks
Anthropic CEO Dario Amodei (Krisztian Bocsi/Bloomberg via Getty Images)
Anthropic announced Tuesday that the U.S. government has lifted export controls on its Claude Fable 5 and Mythos 5 models, ending an 18-day shutdown that had blocked all users worldwide from accessing the two models. Starting Wednesday, users around the world will be able to access Fable 5 through Claude.ai, Claude Code, Claude Cowork, and the Claude platform.
Through July 7, subscribers on Pro, Max, Team, and select Enterprise plans will receive Fable 5 access covering as much as half their weekly usage allotment before credits apply, Anthropic said. Reinstatement of Fable 5 on Amazon $AMZN Web Services, Google $GOOGL Cloud, and Microsoft $MSFT Foundry is also planned, though Anthropic did not specify a timeline. A limited group of U.S. organizations can once again reach Mythos 5 after the government gave its sign-off on June 26; extending that access to additional domestic and overseas partners via the Glasswing program remains an ongoing conversation with federal authorities.
Commerce Secretary Howard Lutnick said on X $TWTR that his agency had spent the preceding two weeks vetting Fable 5 alongside Anthropic to achieve alignment across the U.S. government and bolster American leadership in the technology. "Our shared priority remains: get the best tech deployed as quickly and safely as possible," White House chief of staff Susie Wiles said on Xtt
The directive, handed down June 12, stemmed from Amazon researchers' documentation of a technique for eliciting dangerous outputs from Fable 5 related to software vulnerability discovery. Faced with a prohibition covering every foreign national — a category that included researchers employed at Anthropic itself — the company concluded the only compliant path was to take the models offline entirely. Anthropic said its own testing found that other widely available models, including OpenAI's GPT-5.5, could produce the same outputs flagged in the report.
Anthropic developed a new safety classifier in response; in testing, it neutralizes the Amazon-documented bypass technique in over 99 percent of attempts, the company said. The Commerce Department's Center for AI Standards and Innovation evaluated both the original guardrails and the newly added classifier, Anthropic said, with results confirming their effectiveness. Blocked requests to Fable 5 will be redirected to Claude Opus 4.8.
The shutdown added disruption to Anthropic at a delicate moment, with the company having filed a confidential IPO prospectus disclosing a revenue run rate of $47 billion and a valuation of $965 billion. According to CNBC, co-founder Tom Brown stepped into the primary negotiating role in place of CEO Dario Amodei — a dynamic reflected in Lutnick's correspondence, which was directed to Brown.
More broadly, Anthropic said the incident revealed that the AI industry lacks a consistent approach to evaluating and responding to jailbreaks, the term for exploits that circumvent a model's built-in safety constraints. To close that gap, Anthropic said it is collaborating with Amazon, Microsoft, Google, and additional Glasswing partners on an industry-wide rubric that would score jailbreak severity along four axes: capability gain, breadth of capability gain, ease of weaponization, and discoverability. The company said it would also launch a HackerOne program for security researchers to submit potential cyber jailbreaks in Fable 5 for review. Anthropic added that it will provide designated government partners with expanded early access to future models that materially advance capabilities in areas relevant to national security.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
