Anthropic pulls top AI models Fable 5 and Mythos 5 worldwide after US export ban

Amazon's reported role in the decision

According to US media reports by The Wall Street Journal, Reuters, and The Information, Amazon CEO Andrew Jassy was one of the driving forces behind the export ban. Jassy told senior US government officials, including Treasury Secretary Scott Bessent, that Amazon's cybersecurity researchers had successfully 'jailbroken' the Fable 5 model using a targeted prompt sequence. In a jailbreak, users find ways to bypass the safety guardrails built into an AI system. The method essentially involves instructing a model to analyze a software codebase and identify or fix possible errors. Jassy said the jailbroken model then produced sensitive information that could be used for large-scale cyberattacks. Amazon is one of Anthropic's largest investors and is closely linked to the AI company through its cloud division AWS.

Tech billionaire David Sacks, described as a former 'AI Czar' of the Trump administration and co-chair of the President's Council of Advisors on Science and Technology, confirmed the role of an unnamed 'partner' — understood to be Amazon — that presented a functioning jailbreak to the government. "Die Administration forderte [Anthropic-CEO Dario Amodei] auf, den Jailbreak entweder zu flicken oder das Modell vorerst vom Netz zu nehmen. Dario weigerte sich," Sacks said, according to finanzen.net's translation. He added: "Ein hochgradig glaubwürdiger, vertrauenswürdiger Partner sowohl von Anthropic als auch der US-Regierung trat mit einem funktionierenden Jailbreak an uns heran." Sacks said that after Anthropic's refusal, the government "widerwillig" pulled the trigger and imposed the export ban. He also expressed the hope that Anthropic would fix the security issue, after which the export controls would be lifted and Fable could return to general release.

Anthropic's response and the proportionality dispute

An Amazon spokesperson declined to share details of those discussions, telling finanzen.net: "Wenn dies geschieht, teilen wir die Details dieser Diskussionen nicht." The spokesperson added that, as a leading cloud provider serving a large number of private and public sector customers, it is not unusual for governments to seek the company's advice on potential security risks.

Anthropic acknowledged that the authorities had learned of a method that could partially bypass Fable 5's protective mechanisms, but said it had reviewed a demonstration of the technique and found that it exploited a few already known and relatively simple vulnerabilities. The company stressed that no universal jailbreak has yet been discovered — that is, no method that would comprehensively disable protective measures and unlock numerous cyber capabilities. It also noted that the agency has so far provided only verbal indications of a possible, narrowly limited, non-universal jailbreak, and that comprehensive protection against all jailbreak techniques is currently not possible across the industry.

In a blog post, Anthropic countered that the government had only received information about a 'potential,' not a verified, critical jailbreak, and questioned the proportionality of the measure. "Anthropic entschuldigte sich bei seinen Kunden für die Einschränkungen und erklärte, man gehe derzeit von einem Missverständnis aus," finanzen.net reported. The company said Fable 5 is already used by hundreds of millions of people and that the restrictions impose a significant burden on its commercial operations. It also noted that Mythos had been marketed for weeks as 'too dangerous for the public' in order to generate exclusivity — a positioning that contrasts with the company's current insistence that the models are safe for general use.

What happens next

Anthropic said it is working to restore access to Fable 5 and Mythos 5 as quickly as possible and stressed that other Anthropic models are not affected by the measure. The company also argued that authorities should have the ability to prohibit the release of unsafe AI models, but said such a procedure must be transparent, fair, and technically traceable. According to Sacks, the administration's immediate hope is that Anthropic addresses the security issue so that the export controls can be lifted and Fable can return to general release.

The dispute highlights the unusual entanglement between Anthropic and Amazon, which is both a major investor in the AI company and a competitor in cloud services through AWS. The episode also raises broader questions about how governments should weigh national security concerns against the commercial availability of frontier AI systems, particularly when competing models offer comparable capabilities. With Anthropic working to restore access and the administration signaling it wants the security issue fixed rather than the models permanently removed, the immediate question is how quickly the company can close the specific vulnerability and resume general release.