Export-Control Order Forces Shutdown of Fable 5, Mythos 5
In a significant move, the U.S. government has mandated the shutdown of two advanced artificial intelligence models, Fable 5 and Mythos 5, developed by Anthropic. This decision, enacted under an export-control directive, prohibits access to these technologies for foreign nationals, compelling the company to disable the models just three days after they were publicly released. The compliance was marked by resistance from Anthropic, which expressed concerns over the lack of a clear rationale from the government.
Commerce Secretary Howard Lutnick took the initiative by sending a letter to Anthropic’s CEO, Dario Amodei, enforcing broad restrictions that extend to foreign nationals both inside and outside the United States, which notably includes Anthropic’s own employees. Given the complexities involved in verifying the nationality of each user in real time, Anthropic has chosen to disable access to both models for all customers to ensure adherence to the directive.
In a perplexing turn of events, Anthropic communicated that the government letter did not provide specific details about the national security concerns prompting this sudden action. The government’s communication merely included verbal evidence regarding a potential jailbreak of the Fable 5 model.
Fable 5 was initially launched as the first publicly available model in Anthropic’s Mythos class, designed to be the company’s most capable offering to date. The Mythos 5 model, while built on similar underlying technology as Fable 5, was kept restricted to Anthropic’s Project Glasswing. This project specifically caters to vetted cyber defenders and infrastructure providers.
Adding another layer to the situation, Mozilla reported on the efficacy of earlier Mythos models, asserting they had resolved numerous vulnerabilities. Both Fable 5 and Mythos 5 were priced for commercial use, set at $10 per million input tokens and $50 per million output tokens, reflecting a significant investment for users looking to leverage these cutting-edge technologies.
The Fable 5 model was built with robust safeguards aimed at facilitating safe general release. It employs a system of classifiers designed to identify potential misuse before responding to user queries. When queries involve complex subjects such as cybersecurity, biology, chemistry, or model distillation, the Fable 5 routes them to a less capable yet safer model known as Claude Opus 4.8. Impressively, over 95% of sessions involving Fable did not trigger any fallback, indicating a strong effectiveness of the model’s protective measures.
Prior to its launch, Anthropic indicated that it had collaborated extensively with U.S. government representatives, the United Kingdom’s AI Safety Institute, and various private organizations to rigorously stress-test Fable’s safety features. Testing reportedly showed no testers were able to discover a universal jailbreak capable of overcoming the model’s protections across a spectrum of tasks.
Moreover, Anthropic established a new data retention policy that obliges the company to keep all traffic associated with Fable 5 and Mythos 5 for at least 30 days. This measure aims to support ongoing research and mitigation efforts concerning potential jailbreak vulnerabilities. The data collected will not be utilized for model training and will be deleted after 30 days, as per the outlined policies.
While analyzing the foundation of the government’s directive, Anthropic highlighted that the capabilities demonstrated in their report could similarly be found in other publicly available models, notably OpenAI’s GPT-5.5, which remains unaffected by such export controls. The company openly disagreed with the government’s rationale, suggesting that identifying a narrow potential jailbreak should not warrant the drastic measure of recalling a commercially viable model that has been deployed widely to millions.
The situation escalated further with reports indicating that the previous Trump administration had also attempted to persuade Anthropic to postpone the release of Fable 5 before the government’s formal directive was issued. This friction gained attention earlier in the year when negotiations between Anthropic and the Department of Defense fell through. Subsequently, the DOD had classified the AI firm as a supply chain risk, a classification that Anthropic has since challenged in court.
The recent export-control order has incited a strong backlash from various stakeholders within the industry. Dean Ball, an AI policy expert who briefly served in the Trump administration, publicly condemned the decision, labeling the move as "simply cartoonish" on social media platform X, previously known as Twitter.
Cybersecurity researcher Peter Girnus posited that Anthropic’s own marketing language, which frequently referred to its products as unique munitions, fostered a climate where the government felt justified in its actions. He stated that the company essentially authored the legal basis for this outcome and marketed it as a brand.
The workforce implications of the directive have raised additional concerns. Gaurav Aggarwal, a technology professional, posed a critical question on LinkedIn regarding whether foreign nationals working at American AI companies are now effectively barred from utilizing the very models they have contributed to creating. This could potentially lead to scenarios where these technologies are deployed against the countries of citizenship of those employees.
In conclusion, Anthropic has stressed that applying such stringent standards across the industry could halt all new model deployments for the majority of leading AI providers. The company advocates for a statutory framework that allows the government to impede unsafe deployments but emphasizes that this specific action fails to adhere to principles of transparency, fairness, and technical accuracy. As the implementation of this directive unfolds, clients who had received complimentary access to Fable 5 until June 22 have also lost that privilege, leaving many in the industry to grapple with the ramifications of this unexpected shutdown.