Claude Fable 5 and Mythos 5: US lifts export controls

The Trump administration has lifted export restrictions on Anthropic’s Fable 5 and Mythos 5 AI models, ending an 18-day disruption that blocked access for users worldwide. Sophia Cai and Cheyenne Haslett report for Politico that Commerce Secretary Howard Lutnick informed Anthropic it no longer requires an export license, provided the company proactively detects security risks, coordinates with the government on future releases, and reports any malicious activity found in its models.

Fable 5, the consumer-facing version of Mythos 5, becomes available again on today, July 1, across Claude.ai, Claude Code, and other Anthropic platforms. Access to Mythos 5 had already been partially restored for a select group of US organizations on June 26.

The controls were triggered on June 12 after Amazon researchers found a method to bypass Fable 5’s safeguards, prompting the model to identify software vulnerabilities and, in one instance, produce code demonstrating how to exploit one. Anthropic says its own testing showed that multiple less capable models, including GPT-5.5 and Claude Opus 4.8, could produce the same outputs, suggesting no unique dangerous capability had been unlocked.

A new framework for AI jailbreaks

Anthropic has since deployed an improved safety classifier that blocks the reported technique in over 99% of cases. The company acknowledges the fix may also block more legitimate coding and debugging requests, and says it will work to reduce those false positives over time.

Beyond the immediate fix, Anthropic is pushing for broader industry change. Together with Amazon, Microsoft, Google, and other partners, it is developing a shared framework to rate the severity of AI jailbreaks, the prompting techniques that bypass a model’s safety guardrails. The framework would score jailbreaks on four criteria:

• Capability gain: Does the jailbreak unlock abilities beyond what existing tools can already do?
• Breadth: Does the same technique work across many different harmful tasks?
• Ease of weaponization: How much effort does it take to turn the jailbreak into an actual attack?
• Discoverability: Is the technique already widely known or hard to find?

Anthropic also committed to giving government partners early access to new frontier models for independent evaluation before broad release, and to sharing threat intelligence with relevant agencies. The company says it hopes these steps will eventually form the basis for formal industry-wide regulation.

Sources

• Redeploying Claude Fable 5 – Anthropic
• Trump lifts limits on Anthropic’s Fable and Mythos models – Politico