Nvidia and Palantir Restrict Anthropic's Fable Model
Major enterprise clients are restricting their use of Anthropic's Fable model over data retention policies, highlighting a growing trust gap between AI developers and corporate customers.
Enterprise adoption of Anthropic's flagship Fable model has hit a roadblock as major clients, including Nvidia and Booz Allen Hamilton, restrict its use for sensitive operations. The backlash stems from a policy update Anthropic introduced in June, which allows the lab to retain Fable usage logs for 30 days to protect against security threats. In response, Nvidia has relegated Fable to non-sensitive open-source projects, choosing to run its own Nemotron models for internal tasks like supply chain monitoring. Similarly, Booz Allen Hamilton, which previously utilized Anthropic's Mythos model, has barred its staff from using Fable to develop proprietary cybersecurity software.
Palantir is also pushing back, blocking Fable deployments for its clients until Anthropic provides permanent zero-data-retention guarantees. To appease corporate users, AI labs are adjusting their offerings. Following OpenAI's decision in August to let GPT-5.6 Cyber users store security logs on local servers, Anthropic plans to launch a comparable program for select clients this fall. However, even under zero-data-retention agreements, providers like OpenAI and Anthropic continue to gather de-identified metadata, raising concerns among practitioners about intellectual property exposure.
Industry experts warn that de-identification offers weak protection. John Schulman, an OpenAI co-founder now at Thinking Machines, noted that labs can still extract intellectual property through reinforcement learning tasks built on user traces. This anxiety was recently highlighted by mathematician Tristan Buckmaster, who questioned if OpenAI's Codex learned from his uploaded Navier-Stokes equations drafts. Although OpenAI concluded that prompts submitted before the September 8, 2026 publication did not influence their system, the incident underscores the fragile trust between researchers and AI developers.
For AI practitioners, this friction signals a shift away from blind reliance on third-party API models for proprietary workflows. Developers working with sensitive code or trade secrets must weigh the convenience of frontier models against the risk of indirect data harvesting. To protect their intellectual property, enterprise teams may increasingly need to deploy self-hosted open-source models or negotiate strict, verifiable zero-data-retention terms before integrating external AI services.
This is our own summary of reporting by The Decoder
