Google DeepMind runs the first double-blind evaluation of a proprietary frontier model, hiding weights and test prompts from each side

Google DeepMind, corroborated by The New Stack and Techmeme·high signalXBlueskyLinkedInCopy link

The pilot puts evaluation inside Confidential Space on Google Cloud's Confidential Computing so that Gemini Flash Lite's weights stay private from the evaluators while the evaluators' prompts stay private from Google. Partners are the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons, and the stated target is benchmark contamination, the problem where a model has already seen the questions. Announced 2026-08-27 with an accompanying technical report; Google frames it as a template for third-party oversight that does not require handing over either data sovereignty or model IP.