Anthropic Remains in Disagreement with the White House Regarding Claude Fable 5

Officials from the Trump administration wrapped up discussions with Anthropic on Monday without lifting the export restrictions imposed last week on the company’s most sophisticated AI models due to concerns about jailbreaking, according to three individuals familiar with the situation.

The administration maintains that there are methods to disable some of the protective measures on Anthropic’s Claude Fable 5, effectively granting users access to the more advanced cybersecurity features of the company’s Mythos model, the sources indicated.

Anthropic has argued for several days that the administration’s concerns are exaggerated, a stance it reiterated during meetings at the Commerce Department with government researchers from the Center for AI Standards and Innovation and the Office of the National Cyber Director, Sean Cairncross, according to one of the individuals.

The meetings also included Commerce Secretary Howard Lutnick, who participated via conference call from the G7 summit in Evian, France. Cairncross himself was not involved, the individual stated.

On behalf of Anthropic, co-founder and Chief Compute Officer Tom Brown, along with Head of External Affairs Sarah Heck, have been steering the conversations. Additionally, Anthropic’s Head of Frontier Red-Teaming, Logan Graham, and Senior Security Researcher Nicholas Carlini traveled to Washington, DC, for the discussions.

“Both parties are working swiftly to resolve this issue,” an Anthropic spokesperson stated in a communication with WIRED. A White House representative declined to provide comments.

It remains unclear what the next steps will entail. The Commerce Department has shown a willingness to find a way to reactivate Fable 5 for consumer use, but it likely hinges on Anthropic fully addressing the jailbreak concerns, the individual noted.

Ringing the Alarm

These urgent discussions have arisen at a tense political time for Anthropic, which is already embroiled in a lengthy dispute with the Pentagon regarding the potential military applications of its AI models.

The Trump administration was initially made aware of the jailbreak concerns last week. Amazon CEO Andy Jassy contacted Treasury Secretary Scott Bessent directly regarding the alleged vulnerabilities, which contributed to alarming the administration, the sources said. Jassy’s dialogue with the Trump administration was first reported by The Information.

In response, anxious White House officials enlisted the NSA to investigate the vulnerabilities. The NSA indicated that it believed it was indeed feasible to remove Fable 5’s protective measures, leading the administration to enact restrictions on the model.

Lutnick subsequently spoke with Anthropic CEO Dario Amodei on Friday as the Commerce Department drafted its letter instituting export controls on Fable 5. Over the weekend, after Anthropic restricted access to the model for all users, Lutnick was in multiple discussions with Brown and Heck, according to a knowledgeable source.

It is unclear why Amazon, one of Anthropic’s major investors, raised alarms about Fable 5. “As a prominent cloud service provider catering to numerous private and public sector clients, it’s common for governments to seek our advice on potential security concerns,” an Amazon spokesperson tells WIRED. “In such instances, we do not disclose the specifics of these discussions.”

Security Disconnect

At the heart of the discussions between Anthropic and the administration is a disagreement regarding the gravity of the Claude Fable 5 jailbreaking concerns.

In a blog post on Friday, Anthropic suggested that the administration’s depictions of the risks are exaggerated. Some cybersecurity experts reiterated this perspective to officials on Monday, sending an open letter asserting that the export control measures taken against Anthropic were unwarranted.

“Anthropic’s Mythos-class models excel at identifying flaws and exploiting vulnerabilities. However, they are not uniquely proficient at these tasks, and many of the undersigned regularly utilize other foundation and open-source models for security assessments and red-teaming every day,” the open letter states. “Consequently, this action has deprived defenders of the best models, created market instability, and jeopardized America’s AI leadership without any legitimate risk to warrant it.”