Confidential Space

Confidential Space is like a locked box of a computing environment: two different people can each put their own secret stuff in, and the box does the computation for them without either one seeing what the other put in. Each side drops their own material into the box, the box mixes it internally, and only a single result comes out. Nobody holds the key to open the box, so not even the people who put things in can see what the other party contributed.

This box turns out to be very handy for evaluating AI models. If an evaluation body just hands its test questions directly to a model company, the model could end up trained on those questions beforehand, inflating its score. On the other hand, if the model company hands over the model itself, there's a risk the technology leaks out. With Confidential Space, the evaluation body never sees the model and the model company never sees the questions — the test runs entirely inside the box, and only the scored result comes out.

Google DeepMind said it used Confidential Space, part of Google Cloud's confidential computing portfolio, for this pilot evaluation. Inside this environment, the evaluators couldn't see the model weights, and Google couldn't see the test questions. A common misconception is that this isn't the name of some specific AI model — it's a Google Cloud infrastructure service that performs computation on behalf of both sides while keeping each party's secrets hidden from the other.