Nvidia wary of Anthropic's Fable for sensitive work

The friction runs through Anthropic's own backers and customers. Nvidia has invested in Anthropic, reportedly plans to continue doing so, and supplies the company with hardware for model development, yet it uses Fable only for less sensitive tasks. Booz Allen Hamilton, one of the earliest users of Anthropic's Mythos model, has banned employees from using Fable for work on proprietary cybersecurity software, with CTO Bill Vass saying the company worries Fable might be learning from some of its code.

Anthropic has tried to respond. After customer pushback and OpenAI's August move to let GPT-5.6 Cyber customers store security logs on their own servers, Anthropic followed suit with a similar program rolling out to select customers this fall. But zero data retention still leaves gaps, because both OpenAI and Anthropic collect metadata and technical usage data from enterprise customers, according to The Information.

OpenAI co-founder John Schulman and AI researcher Sarah Hooker point to deeper limits. Schulman laid out a spectrum of training approaches that can still extract customer IP, called de-identification weak, and said users can be traced back with just a small number of bits. Buckmaster's Navier-Stokes dispute with OpenAI, even after OpenAI's internal investigation cleared the Codex prompts of influencing the system, showed how fragile the trust between business, academia, and the AI labs really is.