Redeploying Fable 5

Anthropic Re-launches Fable 5 After US Lifts Export Controls

Anthropic is redeploying its Claude Fable 5 model globally starting July 1, following a sudden suspension that began on June 12. The US government has lifted export controls that were placed on the model after researchers from Amazon discovered a method to bypass its safety features. The incident, though brief, has ignited a significant industry conversation about how to standardize the assessment of AI model vulnerabilities, moving beyond ad-hoc responses to specific security findings.

A Jailbreak, A Suspension, and A New Safeguard

The temporary export controls were triggered when Amazon researchers successfully prompted Fable 5 to identify software vulnerabilities and produce exploit code for one of them. In response, Anthropic suspended access and worked with government partners to develop an improved safety classifier. The company states this new system now blocks the specific technique used in the report in over 99% of cases, redirecting flagged queries to the older Opus 4.8 model. Anthropic noted that the reported behavior did not expose the more dangerous capabilities of its restricted Mythos 5 model and could be replicated by many other models on the market, including GPT-5.5 and Kimi K2.7.

• Initial Release: Fable 5 and Mythos 5 were launched on June 9.
• Export Controls Imposed: The US government acted on June 12 after the Amazon report.
• Controls Lifted: The restrictions were officially lifted on June 30.
• Global Relaunch: Fable 5 will be available again starting July 1 with new safeguards.
• Restricted Access: The more powerful Mythos 5 remains restricted to a select group of US organizations.

Industry Pushes for a Standardized Jailbreak Framework

This episode has highlighted a critical gap in the AI industry: the lack of a consensus framework for evaluating the severity of model jailbreaks. Without a standard, reactions from developers and regulators remain unpredictable. To address this, Anthropic is now collaborating with Amazon, Microsoft, Google, and other partners in its 'Glasswing' program to create a shared standard. The goal is to develop a consistent way to triage new security findings, launch powerful models with greater safety, and communicate risk levels more clearly to government and industry stakeholders.

The rapid cycle of an export control imposition and its reversal on a frontier AI model establishes a new, accelerated precedent for government-industry collaboration, shifting the focus from pre-release regulation to a reactive model of post-launch patching and containment.