Microsoft’s Data Retention Policy: Claude Fable Restricted for Employees

Microsoft’s Concerns Over Anthropic’s Claude Fable 5 AI Model

Anthropic recently unveiled Claude Fable, its inaugural Mythos-class AI model, which has already stirred up apprehension within Microsoft. Sources reveal that Microsoft has imposed restrictions on the utilization of Claude Fable 5 by its employees due to Anthropic’s updated data retention policies.

Despite the swift integration of Claude Fable 5 into Microsoft’s GitHub Copilot and Foundry services for external customers, the model remains inaccessible through the internal model picker used by Microsoft employees for GitHub Copilot versions. This disparity stems from Anthropic’s requirement of data retention for Claude Fable 5, contrasting with the Zero Data Retention (ZDR) protocol adhered to by other Claude models within Microsoft.

Microsoft is reportedly evaluating the implications of Anthropic’s data retention policies, particularly concerning customer data and confidential information. The outcome of these legal deliberations will determine whether Claude Fable 5 receives approval for internal use at Microsoft.

Claude Fable 5 relies on data retention to support Anthropic’s safety classifiers, necessitating the retention of prompts and outputs for a specified duration before deletion. Legal uncertainties arise from the storage of flagged prompts and outputs for up to two years if they violate Anthropic’s usage policy, prompting caution in Microsoft’s utilization of Claude Fable 5.

In response to inquiries, Microsoft has declined to provide comments on the matter.

The introduction of Claude Fable 5 marks a significant milestone in Anthropic’s Mythos AI model lineup, following previous statements cautioning against the release of the family for cybersecurity purposes. Anthropic has implemented prompt safeguards to mitigate risks associated with Fable 5, prompting the adoption of new data retention measures.

Update, June 10th: It is noted that Microsoft has refrained from offering further comments on this issue.