On Wednesday, the U.S. government lifted export controls on top-tier artificial intelligence (AI) models classified as Mythos. Anthropic announced plans to enhance safety measures and resume distribution of its most advanced AI models starting tomorrow. Consequently, users in South Korea are expected to regain normal access to Claude Mythos 5 and Claude Fable 5, which had been restricted until now.
Anthropic reported via its X (formerly Twitter) account on Tuesday that it had received notification from the U.S. Department of Commerce lifting export controls on Claude Fable 5 and Mythos 5. The company stated that it will begin restoring access tomorrow and share updates soon.
Later, Anthropic announced on its website that Fable 5 will be available globally starting tomorrow through the Claude Platform, Claude.ai, Claude Code, and Claude Co-Work.
However, direct access to Mythos 5 is expected to take longer to restore.
Anthropic disclosed that they have restored access to Mythos 5 for select U.S. institutions and are in ongoing negotiations with the government to expand access for domestic and international partners involved in Project Glasswing.
Export Controls Lifted After 18 Days; We Appreciate Users’ Patience
On June 12, the U.S. government had issued export control guidelines that completely halted foreign access to Mythos 5 and Fable 5. Both are Anthropic’s top-tier AI models, with Mythos access limited due to cybersecurity concerns. Fable 5 is a publicly available model equipped with safety measures.
On June 26, the U.S. Department of Commerce allowed ‘Claude Mythos 5’ to be provided only to certain verified U.S. institutions and users. This relaxed the previous restrictions on foreign access to Mythos but still applied only to domestic entities, leaving South Korea’s access restricted.
The South Korean government and several domestic companies had joined Project Glasswing to gain access to Mythos, but U.S. government actions had stalled the model’s utilization. Participants included government agencies like the Korea Internet & Security Agency (KISA) and companies such as Samsung Electronics, SK Hynix, and SK Telecom.
With the lifting of these export controls, South Korea’s access to Mythos 5 and Fable 5 is expected to be restored soon. The Ministry of Science and Information and CommunicationsTechnology (ICT) has stated that they are in ongoing communication with Anthropic.
Anthropic expressed gratitude for users’ patience and thanked everyone who collaborated on the model’s redistribution.
Amazon Highlights Palisade 5 Jailbreak Issue; Additional Safety Measures Implemented
The export control situation began when Amazon reported potential jailbreak vulnerabilities in Fable 5 to the U.S. government, raising concerns about bypassing safety measures for cyberattacks.
In its announcement, Anthropic acknowledged the vulnerabilities reported by Amazon. However, they emphasized that the reported techniques do not expose special cyberattack capabilities at the Mythos level. The company stated that they have enhanced safety measures to block the reported jailbreak techniques.
Anthropic also announced plans to continue collaborating with the U.S. government on advanced AI security. For AI models directly related to national security, they will provide pre-access to the government and cooperate in safety assessments before release. They plan to promptly share cybersecurity vulnerability information with government agencies if any bypasses or misuses of AI model safety measures are detected. The company will also significantly expand joint research on AI security with the government.
Anthropic Launches Cost-Effective Claude Sonnet 5
Alongside the news of restoring access to Fable 5, Anthropic announced the launch of its cost-effective AI model Claude Sonnet 5.
This new model offers performance close to Opus, which ranks just below Mythos in the AI hierarchy, but at a lower price point. It significantly improves on key capabilities such as reasoning, tool usage, coding, and knowledge tasks compared to its predecessor, Claude Sonnet 4.6.
Notably, Sonnet 5 can complete complex tasks that previous Sonnet models would have abandoned mid-process. It can also independently verify its results without additional prompts. Safety improvements have been made, with reduced rates of hallucinations and unwarranted flattery.
Starting today, Sonnet 5 will be available as a basic model in both free and pro plans.