Anthropic has announced the global re-deployment of Claude Fable 5, making the AI model available to users worldwide while introducing enhanced transparency around its cybersecurity safeguards and a proposed framework for assessing the severity of AI jailbreaks.

Alongside the relaunch, the company has shared additional details about the AI-powered cybersecurity classifiers that operate alongside Claude Fable 5. These classifiers are designed to detect and block dangerous or potentially harmful cybersecurity-related requests, helping prevent the model from being used for malicious activities while enabling legitimate use cases.

According to Anthropic, the newly published documentation outlines the specific categories of cybersecurity harms the classifiers are designed to prevent, as well as the types of activities that fall outside their scope. The move is intended to provide greater clarity to users, researchers, and policymakers about the model's safety capabilities and limitations.

In a parallel effort to improve AI safety governance, Anthropic has also released an early draft of its proposed AI Jailbreak Severity Framework, developed in collaboration with its Glasswing partners.

AI jailbreaks refer to techniques that manipulate prompts to bypass a model's built-in safety safeguards, potentially enabling responses that would otherwise be restricted. Anthropic noted that jailbreaks vary significantly in impact—some may only circumvent minor restrictions, while others can expose a broad range of harmful capabilities, including dangerous cybersecurity-related outputs.

The proposed framework seeks to establish a common language for classifying the severity of jailbreaks based on the extent of safety controls they bypass and the associated risks. According to the company, the absence of a standardised methodology has made it difficult for AI developers, regulators, and researchers to consistently assess and communicate the significance of jailbreak vulnerabilities.

By introducing the draft framework, Anthropic aims to facilitate more effective collaboration between AI developers, governments, and the broader AI safety community. A standardised approach to evaluating jailbreak severity could support coordinated vulnerability reporting, risk assessment, and the development of proportionate mitigation strategies as advanced AI systems become more widely deployed.

The announcement reflects the growing industry focus on responsible AI deployment, particularly as organisations face increasing expectations to demonstrate robust governance, transparency, and security measures for foundation models used in sensitive domains.