AI company Anthropic revealed that it received an export control directive from the US government, instructing the suspension of access to its advanced “Fable 5 and Mythos 5” models for foreign nationals, including its employees worldwide. To comply with the directive, Anthropic is promptly disabling these models for all customers, as stated in a blog post by the company. Anthropic expressed disagreement with the government’s decision to recall a commercial model due to a potential jailbreak, arguing that such actions could impede new model deployments across the industry.
The US government, without specifying the national security concern, informed Anthropic about a method to bypass or “jailbreak” Fable 5. Anthropic acknowledged observing a demonstration of this technique revealing minor vulnerabilities, which were deemed simple and detectable by other publicly-available models. Prior to the Fable launch, Anthropic extensively collaborated with various entities to assess the model’s safeguards, asserting that these measures surpass the effectiveness of previously deployed models. The company emphasized that no universal jailbreak method has been identified, capable of broadly circumventing Fable’s safeguards and enabling diverse cyber capabilities.
