Anthropic abruptly disabled access to its Fable 5 and Mythos 5 models after receiving a US government directive at 5:21pm (ET) to halt access by foreign nationals, a move that underscores escalating concerns around artificial intelligence security. Despite rigorous testing with US and UK government agencies, a method of bypassing Fable 5’s safeguards was demonstrated. Anthropic asserts these vulnerabilities are not unique to Fable 5, as other publicly-available models can discover them without requiring a bypass. The company maintains its safeguards are more effective than those of previous models, even though the government’s action highlights the difficulty of fully securing even heavily red-teamed AI systems.
US Government Directive Suspends Fable 5 and Mythos 5 Access
This action, citing national security authorities, underscores the escalating scrutiny of advanced artificial intelligence systems and their potential misuse. Anthropic asserts the vulnerabilities are not unique to Fable 5, despite rigorous testing. The government’s directive did not specify the exact national security concern prompting the suspension, but Anthropic understands it relates to a newly discovered method of bypassing Fable 5’s safeguards. This revelation contrasts with Anthropic’s earlier assertions that its safeguards are more effective than those of previous models and that users had complained they were overly broad. The company maintains its “defense in depth strategy” with Fable 5, aiming to make jailbreaks narrow or expensive, combined with 30-day customer data retention for monitoring and mitigation.
Anthropic disputes the government’s decision to recall a commercially deployed model, stating that if this standard was applied across the industry, it would essentially halt all new model deployments for frontier model providers. The company believes the government should have a transparent, fair, clear, and technically grounded statutory process, and that this action does not adhere to those principles. Anthropic will share more details within the next 24 hours.
In fact, our safeguards are so strong that many users have complained that they are overly broad.
Anthropic
Identified “Jailbreak” Technique & Model Safeguard Effectiveness
Despite rigorous testing with the US government, UK AISI, and third-party organizations, Anthropic encountered a newly discovered method that revealed a small number of previously known, minor vulnerabilities within the model’s safeguards. The government’s action highlights a bypass technique that doesn’t expose a unique flaw in Fable 5, but rather the ability to access vulnerabilities already present in other models. The company emphasizes that the identified vulnerabilities are more effective than those of previous models and are widely available from other models, including OpenAI’s GPT-5.5, which is used daily by cybersecurity professionals. To date, the government has only provided verbal evidence of this potential, narrow, non-universal jailbreak, which Anthropic describes as prompting the model to analyze a codebase and identify software flaws.
This strategy is directly linked to the company’s recent policy change requiring 30-day retention of customer data with Fable, a measure Anthropic admits carries real costs but allows for research and mitigation of potential exploits. GPT-5.5 is used daily by cybersecurity professionals, demonstrating the prevalence of similar vulnerabilities across the industry.
