Trump administration officials concluded negotiations with Anthropic on Monday without lifting export controls that had been in place. imposed last week on the most important aspects of the business advanced AI models in response to jailbreak issues, according to three people briefed on the matter.
The administration continues to believe there are ways to disable some guardrails in Anthropic’s Claude Fable 5, allowing users to access the more powerful cybersecurity capabilities of the company’s Mythos model, the sources said.
Anthropic has said for days that the administration’s concerns were overblown, a position it reiterated in working group meetings held at the Commerce Department with government researchers from the Center for AI Standards and Innovation (CAISI) and the office of National Cybersecurity Director Sean Cairncross, one of the people said.
The meetings were also attended by Commerce Secretary Howard Lutnick, who called from the G7 summit in Evian, France. Cairncross itself did not participate, the person said.
On the Anthropic side, co-founder and CIO Tom Brown and head of external affairs Sarah Heck led the discussions. Anthropic Red Border Team Leader Logan Graham and Senior Security Researcher Nicholas Carlini traveled to Washington, DC for the discussions.
“Both parties are working quickly to resolve this issue,” an Anthropic spokesperson said in a statement to WIRED. A White House spokesperson declined to comment.
It was not immediately clear how the next steps might unfold. The Commerce Department has expressed willingness to find a way to bring Fable 5 back online for consumer use, but that would likely depend on Anthropic fully fixing the jailbreak issues, the person said.
Sound the alarm
The emergency talks come at a difficult political time for Anthropic, which already finds itself in a prolonged period of fight with the Pentagon on whether its AI models could be used for certain military applications.
The Trump administration was first alerted last week to problems with jailbreaking. Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent directly about the alleged vulnerabilities, which played a role in scaring the administration, the sources said. Jassy’s conversation with the Trump administration was the first reported by The Information.
Alarmed White House officials tasked the NSA with helping examine the vulnerabilities. The NSA responded that it believed it was indeed possible to remove Fable 5’s guardrails, prompting the administration to place restrictions on the model.
Lutnick then spoke with Anthropic Chief Executive Dario Amodei on Friday as the Commerce Department drafted its letter imposing export controls on Fable 5. Over the weekend, after Anthropic cut off access to the model to all users, Lutnick had several calls with Brown and Heck, according to a person with knowledge of the events.
It’s unclear why Amazon, one of Anthropic’s largest investors, sounded the alarm on Fable 5. “As a leading cloud provider that serves a large number of private and public sector customers, it is not uncommon for governments to seek our advice on potential security risks,” an Amazon spokesperson told WIRED. “When they arise, we do not share the details of these discussions.”
Security Logout
At the heart of the conversations between Anthropic and the administration is a disagreement over the severity of Claude’s Fable 5 jailbreak issues.
In a blog post On Friday, Anthropic suggested that the administration’s descriptions of potential risks were exaggerated. Some cybersecurity researchers reiterated that position to officials Monday, sending a open letter arguing that the export control measures taken against Anthropic were unjustified.
“Anthropic’s Mythos-class models are quite effective at finding vulnerabilities and exploiting exploits. However, they are not particularly good at these tasks, and many of the undersigned regularly use other foundation and open source models for security auditing and red-teaming every day,” the open letter read. “As a result, this action took away the best models from advocates, created market uncertainty, and jeopardized U.S. leadership in AI without any real risk to justify it.”
Jailbreaking works by tricking an AI model into bypassing its protections in specific ways. Since Fable 5 is a version of Mythos with certain cybersecurity, biology, and chemistry protections, bypassing these protections would effectively give users a version of Mythos. Anthropic itself has expressed serious concerns about allowing Mythos to be used by the general public; However, he said Friday that Fable 5’s safeguards were strong enough to allow for a public release.
Researchers who evaluated Amazon’s findings say the issues identified don’t completely void Fable 5’s warranties. “It wasn’t a jailbreak per se,” says Katie Moussouris, founder and CEO of Luta Security, which published an analysis after reading the Amazon newspaper.
Moussouris points out that even if the US government has proof of a full jailbreak of Fable 5, restricting the model’s ability to access certain topics is a stopgap at best. “Most of us [in security research] I think guardrails are speed bumps and should not be treated as guardrails for experienced opponents,” says Moussouris. “They only serve to slow down the less skilled.”
Anthropic investors also worked this weekend to try to gauge how the company’s latest spat with the White House affects its future, another person close to the company said. Some investors believe the U.S. government is targeting Anthropic and that a competitor might not have faced the same backlash if it had launched a similar model to Mythos, the person said.
The White House export control directive also raises broader questions for other AI labs aiming to release AI models with Mythos-level capabilities, and how they can do so in compliance with the U.S. government. AI labs are now expected to give the White House Early Access to Advanced AI Modelsand be extremely proactive in keeping the U.S. government informed about model launches, according to AI lab executives who spoke to WIRED.
“The weekend events…are informative for everyone who [US] “The government would be willing to take these steps,” says Aidan Gomez, CEO of Cohere, a small Canada-based AI lab that offers enterprise tools. “No one can be naive to this reality.”
