Immediate Halt to Fable 5 and Mythos 5: US Government Blocks Anthropic's Most Powerful Models
The United States government has ordered Anthropic to immediately suspend access to the artificial intelligence models Fable 5 and Mythos 5. This measure, communicated to the company through a directive related to export control regulations, was motivated by alleged national security concerns and resulted in the global deactivation of both models, even for users who would not directly fall within the scope of the restrictions.
According to Anthropic, the order received prevents access to the models for any foreign citizen, both within and outside the United States, including non-US employees of the company. To ensure full compliance with the directive, the company has chosen to completely stop the service for all users. Other models on the platform are reportedly unaffected by the measure.
The decision is particularly significant because it involves the two most advanced systems developed by Anthropic. Mythos 5 represents the company's flagship model and was introduced in limited form at the beginning of April. Anthropic had decided not to distribute it publicly due to its high capabilities in identifying software vulnerabilities. According to the company, during testing, the model was able to identify flaws in all major operating systems and web browsers analyzed. For this reason, Anthropic had launched a controlled program called Project Glasswing, allowing access to about fifty selected organizations active in the cybersecurity sector, including Amazon, Apple, Google, Microsoft, and CrowdStrike. The stated goal was to use the model exclusively for defensive activities and research in the field of cybersecurity.
Fable 5 was launched just a few days ago as a commercially distributable version of the same technology. The model incorporates limitations and protection mechanisms designed to prevent the execution of requests considered high-risk, particularly in the areas of cybersecurity and biological sciences. According to Anthropic, such safeguards made it suitable for public distribution while maintaining much of the original model's capabilities. Some independent benchmarks had already placed it at the top of performance rankings among publicly available AI models.
The origin of the controversy seems to be linked to the discovery of a possible "jailbreak" method, a technique able to circumvent some of the limitations imposed on the model. However, Anthropic claims not to have received detailed evidence from the government, only verbal indications regarding an alleged bypass defined by the company as "limited and not universal." According to the reconstruction provided by the company, the demonstration shown by the authorities would have allowed the model to analyze specific source code and identify some known vulnerabilities of minor significance. Anthropic also argues that similar capabilities are currently available in other publicly accessible models, including OpenAI's GPT-5.5, and that such functionalities are used daily by cybersecurity professionals for defensive purposes.
The company further emphasizes that Fable 5 has undergone thousands of hours of security testing conducted with the involvement of US government agencies, the UK AI Security Institute, independent organizations, and internal teams. From these activities, no universal jailbreak capable of extensively compromising the protections of the model has emerged. Anthropic argues that complete immunity to jailbreaks is currently unattainable for any producer of advanced language models. For this reason, it has adopted a strategy based on multiple levels of protection, continuous monitoring, and data retention for 30 days, a measure introduced to quickly identify and counter any attempts to circumvent safeguards.
While fully respecting the government directive, Anthropic has openly contested the decision of the US authorities. According to the company, the mere finding of a potential limited jailbreak should not be sufficient to justify the withdrawal of a widely distributed model. The company also believes that applying such a criterion to the entire sector could slow down or even block the release of new generations of advanced AI models.