Anthropic vs. Washington: Fable 5 Suspended for National Security

The Fable Model: A Breakthrough Under Surveillance

Anthropic, a company specializing in the development of artificial intelligence models, has recently made headlines with the launch of Fable. This model is a modified version of Mythos, which was deemed too dangerous to be released publicly due to its advanced cybersecurity capabilities. Two months prior to the launch of Fable, Anthropic announced Mythos Preview, a model they considered too risky for public dissemination. Fable, on the other hand, was designed with safety safeguards to mitigate these risks. Despite the skepticism that often surrounds Anthropic's public announcements, Fable has been met with enthusiasm from those who have had the chance to experience it. It stands out for its impressive performance, even surpassing renowned models like GPT 5.5 and Opus 4.8. This advancement recalls the technological leaps observed with previous models such as GPT-4 and Grok 4, suggesting that Fable could well be the precursor to a new generation of AI models.

Conflict with the U.S. Government

However, the launch of Fable has not been without complications. The U.S. government quickly responded by issuing an export control directive, citing national security authorities. This directive prohibits access to Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States, including foreign employees of Anthropic. The company had to disable these models for all its clients to comply with this requirement. Anthropic received this directive from the government at 5:21 PM (ET). The letter did not provide specific details regarding national security concerns. It appears that the government was alerted by a bypass method, or "jailbreaking," which had been reported by Amazon. This situation is particularly notable as Amazon is not only an investor in Anthropic but also a major provider of inference services for the company. It is important to note that access to all other Anthropic models will not be affected by this directive.

Reactions and Implications

Anthropic responded to this directive by emphasizing that non-universal jailbreaks are inevitable and limited, and that there is no evidence of a universal jailbreak. The company examined a demonstration of the reported bypass technique, which allowed for the identification of a small number of minor vulnerabilities that were already known. These vulnerabilities, while present, seem relatively simple and could be discovered by other publicly available models without requiring a bypass. Meanwhile, senior staff members from Anthropic traveled to Washington D.C. to attempt to clarify what they consider a misunderstanding. For their part, officials from the White House suggested that Anthropic's leadership had shown negligence regarding legitimate national security concerns.

The Economic Imperative in the AI Sector

In the field of artificial intelligence, economic value is currently primarily captured by the computing sector. This is due to the fact that supply is failing to meet the growing demand, leading to rising prices. The main beneficiaries of this situation are companies like Nvidia, TSMC, and memory manufacturers such as SK hynix, Samsung, and Micron. In contrast, labs like Anthropic and OpenAI have invested billions in developing cutting-edge models, but these investments are not always profitable. Once released, these models are often distilled and commodified by open-source alternatives, primarily from China. This pessimistic scenario for the labs highlights the risk that they may never cover their costs, as their differentiation is fleeting and free alternatives become "good enough." In such a context, AI models risk becoming commodities, while the majority of the value flows to other sectors, currently computing.

The Importance of Data for the Future of Models

AI models, including Mythos, are not yet at the stage where they can do without additional computing. What they need is more data and better data. Improvements in models increasingly come from reinforcement learning, which can be partially generated synthetically. However, the most powerful leverage for a cutting-edge lab remains real-world usage. This is why OpenAI and Anthropic offer heavily subsidized subscription plans. According to SemiAnalysis, a $200 plan provides access to $8,000 worth of Claude tokens and $14,000 worth of Codex tokens. These companies are fighting not only for the attention of users and developers but also for access to real usage data to improve their models.

Anthropic has intensified its efforts with Fable, announcing that data from all uses would be retained for 30 days, even for enterprise plans that previously promised no data retention. Although the company stated that it would not train on this data, it has not put in place safeguards to prevent future use. If this policy change does not lead to a significant loss of customers, it is likely that Anthropic will eventually use this data, as it is essential for achieving their ultimate goals.