The Trump administration concluded discussions with Anthropic on Monday without lifting export controls on its most advanced AI models, according to three people briefed on the matter. Officials believe the company’s Claude Fable 5 model could be jailbroken, allowing users to access the more powerful cybersecurity capabilities of the Mythos model. Anthropic has consistently denied the administration’s concerns, arguing they are overblown, and reiterated its stance in meetings with government researchers and officials. The talks involved Commerce Secretary Howard Lutnick, who participated via a conference call from the G7 summit in France, and Anthropic’s leadership, including cofounder Tom Brown and head of external affairs Sarah Heck. Anthropic’s spokesperson said both parties are working quickly to resolve the issue, while the White House declined to comment on next steps. The Commerce Department expressed a willingness to bring Fable 5 back online for consumer use, but only if Anthropic fully resolves jailbreak concerns. The situation remains unresolved, with uncertainty about how the dispute will unfold.
The emergency discussions come amid a tense political moment for Anthropic, which has been in a prolonged dispute with the Pentagon over the military use of its AI models. Last week, Amazon CEO Andy Jassy raised concerns about alleged vulnerabilities in Fable 5 with Treasury Secretary Scott Bessent, prompting the administration to act. White House officials then tasked the NSA to review the vulnerabilities, which the agency confirmed could be stripped away, leading to the export controls. Lutnick spoke with Anthropic CEO Dario Amodei on Friday as the Commerce Department prepared its letter. Over the weekend, Lutnick engaged in multiple calls with Anthropic’s leadership after the company cut off access to Fable 5. The reasons behind Amazon’s alarm remain unclear, though the company emphasized its role as a cloud provider and its duty to advise governments on security risks.
The core of the dispute lies in a disagreement over the severity of jailbreaking concerns for Fable 5. In a blog post, Anthropic suggested the administration’s risk assessments were exaggerated, and cybersecurity researchers echoed this view, sending an open letter arguing the export control action was unjustified. The letter claimed that while Mythos-class models are adept at finding flaws, they are not uniquely so, and many researchers use other models for security audits. Researchers who evaluated Amazon’s findings noted the issues did not fully disable Fable 5’s safeguards, with one analyst calling it a stopgap measure rather than a full jailbreak. Investors are also assessing the impact of the dispute on Anthropic’s future, with some suggesting the administration may be singling out the company. The White House’s export control directive raises broader questions for AI labs seeking to release models with Mythos-level capabilities, prompting calls for early government access to advanced AI models.
Source: wired
