On June 12, Anthropic announced that it will "suddenly disable" its most sophisticated AI models, Fable 5 and Mythos 5, for all users. This decision follows a directive from the U.S. government, which mandated a suspension of access for foreign nationals due to perceived national security risks.
The government did not provide specific reasons for its concerns, but suggested that there might be a way to "jailbreak" the Fable 5 model, potentially enabling its use to identify software vulnerabilities. This order arrives as relations between Anthropic and the U.S. government appeared to be improving after previous tensions related to the company’s refusal to allow military applications for its AI technology, which included domestic surveillance and fully autonomous weapon systems. The government subsequently placed Anthropic on a contractor blacklist.
This latest government action signals a significant ramp-up in U.S. efforts to curb potential AI advancements by foreign adversaries, shifting focus from technology hardware to the actual AI models themselves—a notable change in the landscape of U.S. export controls.
Anthropic disagrees with the government's assessment, stating that the agency has only provided "verbal evidence of a potential narrow, non-universal jailbreak." The company argues that a perceived vulnerability shouldn’t warrant such drastic measures, especially for a commercial model that has reached hundreds of millions of users.
Anthropic expressed that their recent call for increased oversight of AI models, along with the ability to restrict those posing unacceptable risks, did not align with the actions taken by the government. Kirsten Davies, the Pentagon's chief information officer, emphasized the importance of national security over commercial interests, reiterating that certain priorities must take precedence.
Anthropic's recent developments occurred shortly after it confidentially filed for a U.S. IPO, potentially positioning itself ahead of competitors like OpenAI. Earlier in the week, the company launched Claude Fable 5, which is classified as a "Mythos-class" model, equipped with safeguards that restrict its use in sensitive domains like cybersecurity. However, some users criticized these safeguards as excessively broad.
Experts warn that the misuse of Mythos models could enhance the capabilities of sophisticated cyberattacks, particularly within industries reliant on complicated legacy technology infrastructures. Anthropic has claimed to have collaborated with governmental entities on safety protocols before the Fable launch, pointing out that similar models from other AI providers exhibit comparable capabilities in identifying coding bugs.
Due to the recent order, Anthropic must deactivate Fable 5 and Mythos 5 for compliance, though access to its other models remains unaffected. The company is hopeful that a misunderstanding has occurred and is working to restore access to the affected models promptly.
In related news, Amazon Web Services confirmed that Anthropic requested the suspension of access to these models for all users across all regions. A U.S. official verified that the Commerce Department's directive prohibits foreign nationals from using Fable 5 and Mythos 5, indicating that non-U.S. citizens, including those residing in the U.S., may need to verify their citizenship to utilize Anthropic's models.
Several prominent figures at Anthropic, such as co-founder Chris Olah and researcher Andrej Karpathy, have origins outside the U.S., though their citizenship statuses were not disclosed. An Anthropic spokesperson refrained from commenting on whether this situation would impact their access to the AI models.
