On June 13, the company Anthropic decided to disable access to its AI models Fable 5 and Mythos 5 for all its clients. This action was taken in response to an order from the US government received on Friday. The developer's other products, including the popular chatbot Claude, continue to function normally.
Reasons for the block and the scope of restrictions
In an official statement, Anthropic clarified that US authorities requested a suspension of access to the new models for all foreign citizens, citing national security concerns. The restrictions are global in nature: they affect users both within the territory of the United States and abroad. Notably, the block even applies to employees of the company itself who are not US citizens.
Although the government has not disclosed the detailed reasons for its decision, insiders at Anthropic believe the incident is related to a jailbreak method for the Fable 5 model. This model was presented to the public on June 9 and was intended to be a more accessible version of the capabilities of Mythos. The latter remains a cutting-edge development in the field of cybersecurity and is available exclusively to partners of the Project Glasswing program.
Technical aspects and vulnerabilities
According to the developers, the capabilities of Fable 5 "exceed" those of any previously released model. In the publication, the company listed the measures taken to ensure security: strict protective mechanisms were implemented to reduce the likelihood of abuse in tasks related to cybersecurity. However, the developers admitted that the protective measures turned out to be so strict that many users complained about their excessive breadth.
The company also noted a fundamental problem in the industry: no provider can ensure perfect resistance to jailbreak attempts, and every model remains vulnerable to specifically tailored attacks. Anthropic's strategy was to make hacks either narrow in scope or extremely expensive to implement, combining this with careful monitoring to quickly stop successful attacks.
Company position and next steps
According to the latest data, the government verbally informed the company about one potential narrow-scope jailbreak. Information about this vulnerability was shared with officials by an unnamed party. Anthropic promised to disclose details of the incident within the next 24 hours.
At the same time, the company stated its disagreement with the notion that a potential jailbreak should be grounds for withdrawing a commercial model. The statement emphasizes that blocking unsafe implementations should occur within a legislative process that is transparent, fair, and based on technical facts. In the opinion of the developers, the authorities' actions do not comply with these principles.