The mystery behind the removal of Anthropic's latest model, Fable5, from the market just three days after its launch has been solved. The swift action, which set a record for the fastest takedown of a cutting-edge AI model, was reportedly instigated by a key stakeholder.
According to multiple foreign media reports, the rare and rapid regulatory intervention by the U.S. government was prompted by Amazon.com Inc. CEO Andy Jassy. Amazon.com is the largest investor and cloud service provider for Anthropic.
During testing of the Fable5 model, Amazon.com researchers discovered that by using a specific series of prompts, it was possible to bypass the model's safety guardrails and access information that could be used for cyberattacks. Jassy personally reported this security vulnerability to U.S. Treasury Secretary Scott Bessent and other senior government officials.
David Sacks, the former U.S. AI lead under President Trump and current co-chair of the President's Council of Advisors on Science and Technology, stated that a "highly credible partner trusted by both Anthropic and the U.S. government" had identified the jailbreak flaw in the model.
Sacks further revealed that the U.S. government presented Anthropic CEO Dario Amodei with a clear ultimatum: either fix the security flaw or take the model offline. Amodei reportedly refused this demand.
Subsequently, the U.S. Department of Commerce issued an emergency export control order to Anthropic on national security grounds. The order mandated the immediate suspension of access to its two newly released flagship AI models, Claude Fable5 and Claude Mythos5, for all foreign nationals.
The scope of the ban was extensive, covering not only all users outside the United States but also foreign citizens within the country. Even foreign employees within Anthropic were prohibited from using these models. As a result, and due to the inability to technically distinguish user nationality with precision, Anthropic was forced to completely disable global access to both models for all users. Other Anthropic models, such as Claude Opus 4.8, Sonnet, and Haiku, remained unaffected.
In an official statement, Anthropic asserted that upon review, the jailbreak method could only identify a small number of previously known minor software vulnerabilities. Furthermore, it claimed these capabilities are already widespread in multiple other publicly available AI models, including OpenAI's GPT-5.5, and are used daily by cybersecurity defenders for system maintenance.
The company warned that if a commercially deployed model used by hundreds of millions of users were to be recalled based solely on the discovery of a narrow, potential jailbreak vulnerability, then, by that standard, all new deployments of cutting-edge AI models across the entire industry would come to a complete halt.
Amazon.com Web Services (AWS) later confirmed in an announcement that it had, at Anthropic's request, revoked all user access to Claude Fable5 and Claude Mythos5 globally. An Amazon.com spokesperson stated in the announcement that it is not uncommon for governments to seek consultation on potential security risks, but the company does not share details of such discussions.
Comments