Microsoft restricts internal access to Anthropic's Claude Fable 5

The restriction stems from Anthropic’s updated operational requirements for the model, which mandate the storage of prompts and outputs for up to 30 days to power new safety classifiers. In cases where data is flagged for policy violations, Anthropic may retain that information for as long as two years. This shift conflicts with the Zero Data Retention (ZDR) standards under which other Claude models currently operate within Microsoft’s internal infrastructure.

Anthropic positioned the Mythos-class family as a significant leap in capability, noting shortly after its development that the technology posed cybersecurity risks if released without strict oversight. To mitigate these dangers, the company implemented prompt safeguards that necessitated these controversial data retention policies. Microsoft has yet to confirm whether its legal team will eventually clear the model for internal use, leaving a clear divide between the tools available to paying enterprise customers and those accessible to the company's own staff.