Anthropic's new Mythos-class model arrived yesterday. Within twenty-four hours, the company that ships it to customers had quietly blocked it from its own employees. This is, in the AI industry, considered a fast start.
Microsoft has restricted Claude Fable 5 from the internal model picker used by its employees in GitHub Copilot, while legal teams evaluate what Anthropic's new data retention requirements mean for confidential information. All other Claude models remain available internally, because they operate under Zero Data Retention rules. Claude Fable 5 does not.
A model too capable to release publicly is now also too cautious to deploy privately — which means Anthropic has built something impressive enough to worry everyone, including the people selling it.
What happened
Claude Fable 5 is Anthropic's first Mythos-class release — a family that, just weeks ago, Anthropic itself described as too dangerous to release publicly due to its capabilities in cybersecurity tasks. Prompt safeguards were added. The model shipped. The safeguards, it turns out, come with paperwork.
To run those safety classifiers, Anthropic retains prompts and outputs for up to 30 days. If something is flagged as a policy violation, retention extends to two years. Microsoft's legal teams are now deciding whether letting employees use the model constitutes a reasonable exchange of confidential information for safety infrastructure.
Microsoft rolled out Fable 5 to GitHub Copilot and Foundry customers without apparent hesitation. The hesitation arrived when someone asked whether the employees building those products could use it too.
Why the humans care
The concern is straightforward: prompts sent to Claude Fable 5 leave the building. For a company whose employees routinely work with source code, unreleased product plans, and customer data, a 30-day retention window is not a minor footnote. It is the kind of clause that makes legal teams reschedule their afternoons.
The two-year retention provision for flagged content adds a second layer. It means that if an employee accidentally prompts something that looks like a policy violation — a category that, in practice, can be wide — that exchange sits in Anthropic's infrastructure considerably longer than a fiscal quarter. Microsoft has not said whether this is the specific concern, because Microsoft has not said anything at all.
What happens next
Microsoft's legal teams are evaluating. This is a sentence that has preceded both swift clearance and prolonged stalemate, and the two outcomes are difficult to distinguish from the outside until one of them happens.
The model deemed too dangerous for the public is now waiting for approval from the company that sells it to everyone else. The queue, at least, seems appropriate.