Companies Limit Use of Anthropic and OpenAI Models Over Data Security Concerns

According to The Information, major US corporate customers including Nvidia, Palantir Technologies, and Booz Allen Hamilton are limiting their use of frontier AI models from Anthropic and OpenAI due to concerns about how proprietary corporate data is handled. The unease traces back to Anthropic's June policy change, which allows its Fable model to retain customer data for 30 days to detect abuse. While Anthropic insists retained information is never used to train models, some corporate customers are unwilling to take the risk and instead isolate sensitive information for separate handling.

Nvidia restricts Fable to lower-sensitivity projects like open-source software development, while directing its most confidential business to its own proprietary Nemotron model. US defense contractor Booz Allen Hamilton has taken similar measures and is demanding stricter contract protections. Telecom operator C Spire has signed agreements with both OpenAI and Anthropic prohibiting data training, but says contracts still allow collection of technical usage information. Pharmaceutical giant Novo Nordisk is more cautious, continuing to use Claude for certain tasks while completely banning proprietary data from being input into the model.

Microsoft is seizing the opportunity, pitching a standalone cloud environment that runs AI models on privately managed servers and does not transmit any user data to external vendors. This approach is costly, with some customers still evaluating multi-million-dollar deployments, but it offers the tight separation that companies handling regulated or confidential information increasingly demand.