MICROSOFT BLOCKS FABLE 5 FOR EMPLOYEES

Microsoft is limiting employee access to Claude Fable 5, Anthropic's newest Mythos-class model, even as it rapidly deploys the same model to paying GitHub Copilot and Foundry customers. The reason: Anthropic's updated data retention policies no longer meet Microsoft's internal security standards for handling confidential corporate information.

Anthropic released Claude Fable as its first Mythos-class model in recent days, a family of AI systems so capable at cybersecurity tasks that the company initially deemed them too dangerous to release publicly. The company has since added prompt-level safeguards to reduce the risk, but the model's architecture still requires data retention to operate Anthropic's new safety classifiers. That requirement is the breaking point for Microsoft's internal use.

Under Anthropic's updated policies, prompts and outputs are deleted after 30 days. However, any content flagged as potentially violating Anthropic's usage policy can be stored for up to two years. For a company like Microsoft, where employees routinely feed proprietary code, internal strategy documents, and customer data into AI assistants, that retention window creates unacceptable risk.

MICROSOFT'S LEGAL TEAMS ARE STILL EVALUATING THE CHANGES

The company has not commented publicly on the assessment. What is clear is the outcome: Claude Fable 5 is not available in the model picker that Microsoft employees use for internal versions of GitHub Copilot. Every other Claude model remains accessible internally because they operate under Zero Data Retention rules, meaning Anthropic does not store any prompts or outputs from those interactions.

The contrast with customer-facing deployment is stark. Microsoft moved quickly to roll out Claude Fable 5 to GitHub Copilot and Foundry customers, where the 30-day deletion window is acceptable for external users making different risk calculations. The company's own employees, however, are being told to stick with models that do not retain their inputs.

The distinction matters because Microsoft's internal AI tools process some of the company's most sensitive data: unreleased product code, internal communications, and confidential business information. Even a 30-day retention window is too long for that class of material, particularly when the vendor can extend storage to two years for flagged content.

MICROSOFT IS NOT MAKING THE SAME BET FOR ITS OWN WORKFORCE

Anthropic's Mythos family was reportedly so capable at cybersecurity tasks that the company originally decided not to release it at all. The capability that made it dangerous, however, is also what makes it useful: strong reasoning about system vulnerabilities, attack patterns, and defensive architecture. Anthropic ultimately decided to release Fable with additional safeguards, betting that the model's usefulness outweighed the residual risk.

Microsoft is yet to comment on record about its internal evaluation or timeline for potentially restoring access. For now, the message to employees is clear: the model is good enough to sell, but not trusted enough to use internally.