At a glance: Anthropic must immediately disable Fable 5 and Mythos 5 following a US government export control directive citing national security concerns; the company disputes that the identified bypass method constitutes a substantive threat.
The US government has instructed Anthropic to block access to the Fable 5 and Mythos 5 models for all users – including foreign employees. The government cites national security concerns regarding a bypass method for the models’ safeguards.
Anthropic received the directive today at 5:21 PM ET. The government letter contained no specific details on the security concerns. According to Anthropic’s understanding, the government obtained knowledge of a method that can bypass Fable 5’s safeguards – a so-called jailbreak. Anthropic reviewed a demonstration of this technique, which revealed only a small number of already-known, minor vulnerabilities. These weaknesses appear relatively straightforward and can also be discovered with other freely available models without bypass methods.
Anthropic bases its security architecture for Fable 5 on multiple layers: the company has implemented strong safeguards that significantly impede misuse for cybersecurity tasks – measures that are reportedly so restrictive that many users have complained. Prior to launch, Anthropic conducted thousands of hours of red-teaming with the US government, the UK’s AISI, and external organizations. These tests demonstrated that Fable’s safeguards are substantially more effective than those of previously deployed models. No tester has found a universal jailbreak that broadly circumvents the safeguards.
Anthropic acknowledges that perfect jailbreak resistance is currently impossible for any model. The company therefore pursues a multi-layered protection strategy: jailbreaks should either be tightly confined or extremely costly to develop, combined with ongoing monitoring for rapid attack detection. This also explains Anthropic’s policy of 30-day data retention for Fable – a measure with significant cost disadvantages, but one that enables jailbreak research.
According to Anthropic’s account, the method obtained by the government essentially consisted of prompting the model to read a specific codebase and fix software vulnerabilities. Anthropic reviewed a report that is said to underlie the government’s decision and validated that the demonstrated capability is widespread among other models – including OpenAI’s GPT-5.5 – and is used daily by security professionals.
Access to all other Anthropic models remains unaffected. Anthropic stated it will publish further details over the next 24 hours.
Source: www.anthropic.com · Published 11 June 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.