Anthropic's newly released Fable 5 model disappeared from public access within days after Amazon alerted the White House to what it characterized as a serious security vulnerability, according to sources familiar with the sequence of events. The incident marks an unusually aggressive intervention in the AI sector and raises sharp questions about how the Trump administration is choosing to regulate cutting-edge artificial intelligence technology.
The timeline moved with striking speed. On Thursday night, Amazon contacted senior administration officials with a report describing how researchers had accessed sensitive portions of Anthropic's more powerful underlying model, called Mythos, through what Amazon characterized as a jailbreak. Within hours, at least five other companies had also flagged concerns to various officials. By Friday evening, the White House issued an export control letter targeting the models. Public access to Fable ended around 10 p.m. that night.
Anthropic had notified the government multiple times in advance of the planned June 9 release and received no objections, according to a source close to the company. The sudden reversal caught the firm off guard. When administration officials called at 1:30 p.m. Friday, they gave Anthropic an ultimatum: take down Fable and Mythos within 90 minutes due to what they called a "national security threat," but provided no specifics about the nature of that threat.
The export control rules imposed go significantly beyond simple geographic restrictions. Rather than merely block foreign adversaries from accessing the models, the controls prevent access by U.S. allies and foreign nationals living in the United States from using them as well. The consequence struck directly at Anthropic's operations: many of the company's workers are foreign-born and require access to these systems for their work.
Anthropic pushed back during subsequent conversations with government officials. The company's leadership, including CEO Dario Amodei, argued that the vulnerability Amazon had identified was relatively basic, achievable with other existing models, and did not represent a fundamental flaw in Fable's safety architecture. The company viewed the government's response as disproportionate to the actual risk.
Katie Moussouris, CEO of Luta Security and someone Anthropic shared the Amazon report with, offered a similar assessment. She noted that the researchers had simply asked the kind of questions a legitimate security defender would pose to test an AI system. Moussouris emphasized that all AI models must be able to assist defensive researchers in exactly this way, or the industry will struggle to scale protections against actual attackers.
"The government's response seems way out of line with what's actually in the research report," Moussouris told Axios.
The intervention carries implications far beyond Anthropic. Industry observers characterized the export control letter as establishing a de facto licensing regime, one in which companies will think carefully before challenging government preferences on AI deployment.
An administration official framed the action differently, saying the government does not view other existing models as national security threats because they have not reached the capability level demonstrated by Mythos. Any model that does reach Mythos-level capabilities would need government approval before release to ensure the nation's defense infrastructure is adequately hardened.
A source familiar with the government's internal thinking suggested that Anthropic bore some responsibility for the escalation, claiming the company had shown a "lack of seriousness" about the release and had been "overly confident" in dismissing the concerns rather than pausing access to address them. Had the company moved more cautiously, the source said, the dramatic Friday night takedown would never have occurred.
The episode also raises murky questions about Amazon's role. As a major investor in Anthropic, Amazon stands to benefit from a regulatory environment that constrains competitors or sets high bars for capability releases. The e-commerce and cloud computing giant declined to detail its conversations with the government, saying only that it is not uncommon for the administration to seek counsel on security risks from leading cloud providers.
Author James Rodriguez: "This reads less like national security policy and more like regulatory muscle-flexing, with Amazon positioned conveniently between the trigger and the target."
Comments