Nvidia, Palantir block Anthropic's Fable over forced data retention policy

新智元 · wechat · 2026-09-15

Anthropic's June 9 policy forces 30-day retention (up to two years for flagged chats) of all prompts and outputs for Mythos-tier models, overriding existing Zero Data Retention agreements with no opt-out.
Fallout: Microsoft restricted internal Fable access; Palantir's Alex Karp publicly attacked AI labs and distributed a client guide on protecting proprietary data, adding OpenAI's GPT-6 Astra instead; Booz Allen runs open models offline; Nvidia swapped supply-chain AI to its own Nemotron.
OpenAI countered on Aug 19 with Private Safety Processing (data stays customer-side); Anthropic responded Sept 1 with Enterprise Frontier Safety alongside Fable5.1 — but it's invite-only, excludes Mythos5.1, and is revocable, failing the "irrevocable" bar.
Ramp data across 70k+ US firms shows Fable5's spend share stuck at 11% while open models' token share jumped from 11% to 62%. Anthropic's own threat-intel report, showcasing its visibility into user conversations, deepened enterprise distrust.

More from Companies & People