The Trump administration’s engagement with Anthropic reached a standstill on Monday as officials concluded discussions without lifting the export controls imposed on the company’s advanced AI models, specifically in relation to concerns over jailbreaking. Sources familiar with the situation revealed that the administration remains apprehensive about the potential for users to disable certain safeguards on Anthropic’s Claude Fable 5, a model they fear could be manipulated to access more sophisticated cybersecurity features from the company’s Mythos model.
Despite the ongoing discussions, Anthropic has repeatedly insisted that the administration’s concerns are exaggerated. This stance was reiterated in meetings at the Department of Commerce, which included government researchers from the Center for AI Standards and Innovation, and involved high-level participation from Treasury Secretary Scott Bessent and Commerce Secretary Howard Lutnick, who participated via conference call from the G7 summit in Evian, France. Notably, Sean Cairncross, the National Cyber Director, did not take part in these discussions.
Leading the talks for Anthropic were co-founder and Chief Compute Officer Tom Brown and Head of External Affairs Sarah Heck. They were accompanied by Logan Graham, Head of Frontier Red-Teaming, and Senior Security Researcher Nicholas Carlini, who traveled to Washington, D.C. For its part, Anthropic issued a statement expressing optimism about reaching a resolution quickly.
The future steps regarding the regulation remain unclear. However, representatives from the Commerce Department indicated a willingness to find a way to reinstate consumer access to Fable 5, contingent upon Anthropic addressing the jailbreak concerns adequately.
These negotiations have unfolded during a particularly tense period for Anthropic, which has been embroiled in a longstanding dispute with the Pentagon about the possible military applications of its AI models. The urgency of the talks intensified after Amazon CEO Andy Jassy alerted Treasury Secretary Scott Bessent to the potential vulnerabilities associated with Fable 5. This direct communication, which was initially reported by The Information, triggered alarm bells within the administration.
In response, the NSA was tasked with assessing the outlined vulnerabilities and reportedly concurred with the potential for disabling the model’s guardrails, leading to the swift imposition of export controls by the Commerce Department. Following this, Lutnick reached out to Anthropic CEO Dario Amodei as the department worked to finalize its regulatory approach. A series of discussions occurred over the weekend as Anthropic took measures to restrict access to the model for all users.
Amazon’s reasons for raising concerns about Fable 5, despite its status as a significant investor in Anthropic, remain ambiguous. An Amazon spokesperson stated that as a major cloud service provider, it often advises governmental entities on security risks, though it refrained from divulging specifics related to this incident.
At the heart of the discussions lies a significant divergence between Anthropic and government officials regarding the seriousness of the jailbreaking vulnerabilities tied to Claude Fable 5. An open letter circulated by several cybersecurity researchers on Monday echoed Anthropic’s perspective, claiming that the administration’s fears were unfounded. The letter pointed out that while Anthropic’s Mythos-class models are proficient at identifying flaws, they are not uniquely capable in this domain, underscoring that other models also serve effectively in security assessments. The researchers contended that by imposing export controls, the government harmed defensive capabilities and jeopardized the U.S.’s competitive edge in AI without any substantial justification for concern.
