Anthropic Takes Claude Fable 5 Offline Under US Government Order
▼ Summary
– Anthropic disabled its Claude Fable 5 and Mythos 5 AI models to comply with a US government export control directive citing national security concerns, issued Friday afternoon.
– The company removed access for all customers, not just foreign nationals, to ensure compliance with the order.
– The incident follows the Trump administration’s earlier designation of Anthropic as a “supply chain risk,” which barred government use of its technology and led Anthropic to file lawsuits.
– Anthropic released Claude Fable 5 on Tuesday with safeguards against questions on cybersecurity, biology, and chemistry, after a limited rollout of Mythos Preview in April.
– The company stated the government’s national security concern is based on a potential narrow jailbreak method, which Anthropic argues reveals minor, previously known vulnerabilities also found in other public models.
Anthropic has confirmed it is taking two newly launched AI models offline after receiving a direct order from the US government on Friday afternoon. The company says it disabled Claude Fable 5 and Mythos 5 to comply with an export control directive citing national security concerns, marking an unprecedented clash between the AI firm and federal regulators.
The order instructed Anthropic to restrict access to “any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” To ensure full compliance, the company has cut off access for all customers, not just those abroad. This incident is the latest in a growing rift between Anthropic and the Trump administration. Earlier this year, the Department of Defense labeled the company a “supply chain risk” after Anthropic attempted to set boundaries on how the US military could deploy its technology. That designation effectively blocked government agencies and contractors from using Anthropic’s systems, prompting the company to file lawsuits against the administration.
Claude Fable 5 was released publicly on Tuesday as a version of the company’s Mythos AI model, equipped with safeguards designed to prevent it from answering questions about sensitive topics like cybersecurity, biology, and chemistry. Anthropic had previously rolled out the Mythos Preview model in April with limited access, aiming to give organizations a chance to harness its powerful cybersecurity tools for defensive purposes while mitigating the risk of misuse by bad actors seeking to create advanced hacking tools.
In a blog post published Friday, Anthropic revealed it received a letter from the US government at 5:21pm ET. “The letter did not provide specific details of its national security concern,” the company wrote. “Our understanding is that the government believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5.” Anthropic added that it reviewed a demonstration of the technique, which was used to identify a small number of previously known, minor vulnerabilities. The company argued these flaws were “relatively simple” and that other publicly available models could discover them without needing a bypass.
The company maintained that it had implemented strong safeguards to reduce the likelihood of misuse and that the jailbreak identified by the government was narrow in scope. “To date, the government has only given us verbal evidence of a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws,” Anthropic stated. It further claimed that an attacker exploiting this method would not be meaningfully more dangerous than one using another AI model.
Representatives for the White House and the US Commerce Department did not respond to requests for comment. Anthropic CEO Dario Amodei, in a policy essay earlier this week, expressed support for a fair, structured, and transparent government process to block unsafe AI releases. However, in Friday’s blog post, the company argued that “this action does not adhere to those principles.”
(Source: Wired)
