Palantir, Nvidia and Booz Allen Hamilton have begun limiting the use of Anthropic's advanced models in work involving sensitive information, amid concerns about data retention and intellectual property. The measures were revealed by The Information on Monday, the 14th.
The restrictions are tied to the security policy adopted by Anthropic for its most advanced models, which calls for data retention for up to 30 days to detect attacks, jailbreaks and other abuse. The company says this information is not used to train new models.
The Palantir requires a zero-retention guarantee before offering Claude Fable directly through its platform. Until that condition is met, customers can still contract the model directly from Anthropic.
The Nvidia uses the model for tasks considered less sensitive, but relies on its own Nemotron models for internal operations involving strategic information. Meanwhile, Booz Allen Hamilton restricted its use in work related to proprietary cybersecurity software developed for clients.
The concern is that code, corporate data or other confidential information remains stored in the vendor's systems, even if only temporarily.
Anthropic prepares alternative for companies
Anthropic says corporate client data is not used for training without authorization. To serve companies with stricter privacy requirements, the company is preparing a solution in which information remains in infrastructure controlled by the client itself.
Meanwhile, eligible enterprise customers can already access versions of the model with zero data retention, reducing one of the main barriers to its use in sensitive corporate environments.
