Export controls lifted, Claude Fable 5 returns with a jailbreak fix that mostly works

Anthropic restored its frontier AI models, Claude Fable 5 and Mythos 5, after the US government lifted export controls that had suspended them since June 12. Starting July 1, the models are available worldwide on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. The suspension was triggered by a now-public jailbreak method, but the re-launch comes with a new classifier that blocks the reported bypass in over 99% of cases. For a breakdown of how these models compare on benchmarks, see the Sonnet 5 technical review.

Timeline of events

The controls were imposed after Amazon researchers found a method to bypass Fable 5's safeguards. The technique let the model identify software vulnerabilities and, in one case, produce exploit code. Anthropic says it worked closely with the government to review the findings. Internal testing showed many less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could produce similar results, raising questions about why only Anthropic's models were suspended. The original suspension and its rationale are documented in the government's initial order.

Anthropic trained an improved safety classifier that blocks the reported bypass in over 99% of cases. Researchers from the US Department of Commerce's Center for AI Standards and Innovation (CAISI) validated the new safeguards. The classifier introduces some false positives during routine coding and debugging, a trade-off Anthropic says it will refine. This kind of granular safety tuning is a theme the latest AI safety framework research also explores.

Fable 5 and Mythos 5 access details

Fable 5 is available on Pro, Max, Team, and select Enterprise plans at no extra cost for up to 50% of weekly usage limits through July 7. After that, access requires usage credits. Access on AWS, Google Cloud, and Microsoft Foundry is being restored as quickly as possible. Mythos 5, released with fewer safeguards for defensive cybersecurity use, was re-enabled for a set of US organizations following government approval on June 26. Anthropic is coordinating with the government to expand access to international Glasswing partners, including its push into the Korean market through the Seoul office expansion.

A shared industry framework for jailbreaks

Together with Amazon, Microsoft, Google, and other Glasswing partners, Anthropic is developing a consensus framework to assess the severity of AI jailbreaks. The proposal scores jailbreaks on four criteria: capability gain, breadth of capability gain, ease of weaponization, and discoverability. The framework is intended to help triage findings, guide responses, and improve communication with government and industry stakeholders. The move mirrors calls for standardized evaluations seen in the TCS-Anthropic partnership for regulated industries.

Deeper government collaboration

Anthropic outlined new commitments: pre-release government access and independent evaluation for models advancing national security capabilities, rapid information sharing on safeguards and threat intelligence, dedicated compute and teams for joint research, and a contribution to a common industry security standard. The company called for these rules to be codified in regulation that applies equally across all frontier model developers. This push for structural rules echoes the argument for specialized AI over generic pilots detailed in the case against enshittification analysis.

"Our hope is that this collaboration, along with our proposed consensus industry framework, will serve as the basis for systematic rules for the whole industry," Anthropic wrote. The company also launched a new HackerOne program for security researchers to submit potential cyber jailbreaks for review. Researchers looking for guidance on testing these systems can consult the best practices guide for AI coding agents.