The U.S. Department of Commerce has officially lifted export restrictions on the Claude Fable 5 and Claude Mythos 5 models, marking a significant milestone in AI technology regulation in 2026. The Anthropic team confirmed this news on June 30, and starting July 1, Fable 5 will once again be available to users worldwide.

The model will be integrated into Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Restoring access through cloud platforms AWS, Google Cloud, and Microsoft Foundry will take a bit longer, but the company promises to do so "as quickly as possible." This decision resulted from productive negotiations with the U.S. government, during which Anthropic agreed to implement a new set of safety classifiers.

New Classifier: Over 99% Accuracy

To resume Fable 5's operations, developers trained a specialized classifier designed to limit behavior described in Amazon's report to authorities. If a user's request is blocked, the system will send a notification and redirect the task to Claude Opus 4.8. According to the company, the new classifier detects the targeted method in over 99% of cases. However, Anthropic acknowledges that this will lead to an increase in false positives during routine programming and debugging tasks—a deliberate compromise for broader access to the model's capabilities.

Mythos 5: Partial Access Already Restored

Access to Claude Mythos 5 was restored for some U.S. organizations on June 26 following government approval. The company continues to collaborate with authorities, including pre-release access to protection tools, sharing information about hacks and abuses, and allocating resources for joint research.

Consensus Framework for Jailbreak Assessment

Anthropic, together with Amazon, Microsoft, Google, and other Project Glasswing partners, has begun developing a unified approach to assessing the danger of jailbreaks. The AI industry still lacks a standard for determining the severity of such attacks, and the proposed scheme evaluates threats based on four criteria:

  • How much the jailbreak expands the attacker's capabilities compared to available tools;
  • How many different malicious tasks the bypass covers;
  • How easily it can be turned into a real attack;
  • How easily others can find or replicate this method.

For the most dangerous cases, the company promises to launch preliminary protective measures immediately after confirming the severity of the threat. Additionally, Anthropic is creating a 24/7 team to monitor channels where jailbreak data is published and launching a HackerOne program for reporting potential vulnerabilities in Fable 5.

My view: Anthropic's decision is a classic example of balancing innovation and safety. Introducing a new classifier with 99% accuracy is an impressive step, but inevitable false positives may frustrate developers. Nevertheless, creating a consensus framework for jailbreak assessment is something the industry has long needed. If Project Glasswing succeeds, it could become a standard for all major AI companies, ultimately strengthening trust in the technology in the long term.