Anthropic has said it will restore public access to Claude Fable 5 and Mythos 5 after U.S. authorities lifted export restrictions that had kept the company’s two most advanced AI models offline since June 12.
Summary
- Anthropic said public access to Claude Fable 5 and Mythos 5 will resume after U.S. authorities lifted export restrictions.
- The models were pulled offline after officials raised concerns over a reported jailbreak that could make Fable 5 identify software vulnerabilities.
- Anthropic said the redeployed models will include new classifiers to block more cybersecurity-related tasks while cooperation with the U.S. government expands.
According to Anthropic, the decision followed “a series of productive conversations” with the U.S. government, after which the company began redeploying the models with new classifiers designed to identify and block more cybersecurity-related tasks.
The company said the latest safeguards are meant to address government concerns linked to possible misuse if the systems are bypassed through jailbreak methods.
The restrictions had forced Anthropic to suspend access to Fable 5 and Mythos 5 for all users earlier this month, after a U.S. government export control directive instructed the company to block both models for all foreign nationals.
In its June 13 statement, Anthropic said the order also covered foreign-national employees working inside the company, prompting it to disable the models entirely to ensure compliance.
U.S. government clears redeployment after review
U.S. Secretary of Commerce Howard Lutnick said on X on Wednesday that officials had worked with Anthropic over the past two weeks to review and approve Fable 5 while keeping the model aligned with U.S. government requirements.
White House Chief of Staff Susie Wiles also said on X that the government’s priority was to get the best AI technology deployed “as quickly and safely as possible.”
The intervention came after officials became aware of a report in which Amazon researchers found a method to bypass Fable 5’s safeguards and make the model identify software vulnerabilities.
Anthropic, however, has argued that the reported issue was not unique to Fable 5, saying weaker models could also identify the same vulnerabilities and produce similar exploit-related output.
In its earlier response to the order, the company said authorities had presented only verbal evidence of what it described as a narrow, non-universal jailbreak. Anthropic said such a method did not remove a model’s safety protections across a wide range of tasks, unlike a universal jailbreak.
“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”
The company had also warned that treating a narrow jailbreak as a reason to recall a commercial frontier model could affect the entire AI industry if applied as a general standard.
Restrictions on Anthropic have raised policy concerns outside the United States as well. On June 29, Austria urged the European Union to explore establishing Anthropic within the bloc, with State Secretary for Digitalization Alexander Proell arguing in a letter that Europe should not risk losing access to major AI advances because of decisions made elsewhere.
Anthropic expands cooperation on AI safety
Alongside the model redeployment, Anthropic said it is increasing cooperation with the U.S. government on model testing, safeguards, and misuse tracking. The company said this will include pre-release access to models and safety systems for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research.
Anthropic has also started drafting a framework with Amazon, Microsoft, Google, and other partners through Project Glasswing, a cybersecurity collaboration announced in April, to assess the severity of AI jailbreaks. The company said the framework is being developed as a consensus effort for classifying jailbreak risks.
The access debate came as Anthropic has continued its push for stricter frontier AI oversight. In its June 11 “Policy on the AI Exponential” proposal, the company called for testing requirements, independent evaluations, cybersecurity standards, and enforcement measures for advanced AI systems, citing potential biological, cybersecurity, and operational risks.
