White House Imposes Export Controls on Anthropic's Fable 5 AI Following Amazon Report

Anthropic's Fable 5 AI model, a general-use version of its more powerful Mythos model, was removed from public access just days after its release on June 9. This swift action followed a report from Amazon that triggered a series of discussions and ultimately, a directive from the White House.

Amazon officials contacted the administration on Thursday night, June 12, to share a report detailing how they were able to "jailbreak" and access parts of Anthropic's Mythos model, which they identified as posing a national security threat. This report, combined with calls from at least five other companies to various senior administration officials, led to the model's shutdown by Friday night.

Anthropic had previously informed the government multiple times about the planned release of Fable 5, and the government had not raised objections, according to a source close to the company. However, administration officials reportedly spoke with Anthropic for hours early Friday in an unsuccessful attempt to persuade the company to withdraw the model.

Around 5:20 p.m. ET on Friday, the White House sent Anthropic a letter, first reported by Axios, stating that the Fable and Mythos models would be subject to sweeping export control rules. By approximately 10 p.m., users lost access to Fable. These controls aim to prevent not only foreign adversaries but also U.S. allies and foreign nationals within the U.S. from accessing Anthropic's most advanced models. This has immediate implications for foreign-born workers within Anthropic who require access to these models.

Anthropic CEO Dario Amodei and other company officials reportedly discussed the situation with the administration after being informed of the impending export rules. During these conversations, Anthropic officials argued that the alleged Amazon jailbreak was relatively simple, could be replicated with other AI models, and did not indicate a flaw in Fable 5's safety systems.

Katie Moussouris, CEO of Luta Security, who reviewed the Amazon report shared by Anthropic, suggested that the government's response seemed disproportionate to the research findings. Moussouris stated that the researchers found vulnerabilities by asking questions typical defenders would pose to AI, which is a designed function of such models for defense purposes.

Administration officials have indicated that other AI models are not currently viewed as national security threats unless they surpass the capability threshold set by Mythos. Any models reaching or exceeding that level would require administration review to ensure the government's national security apparatus is adequately prepared.

According to Axios, a source familiar with the government's perspective suggested a "lack of seriousness" from Anthropic regarding the Fable release, stating that a more proactive response from the company could have prevented the government's intervention.