Tools 2026-08-15

Four working demos of AI running on data it cannot read — and a compiler Google wants ordinary developers to use

Google published a post on 14 August on HEIR, its open-source compiler toolchain for homomorphic encryption — the branch of cryptography that lets a server compute on data it never decrypts. HEIR has been public under Apache-2.0 since 2023; what is new is the pitch and the evidence behind it. Google's stated goal is to make it "a one-click solution to enable non-experts to incorporate encrypted inference into production applications," taking a pre-trained model and converting it to run on encrypted inputs. The post presents four demonstrations measured on a single-threaded CPU: a deep learning recommendation model, credit card fraud detection, network intrusion detection using the Kitsune system, and hotword detection for audio agents. Hardware partners listed include Belfort Labs, Niobium, Cornami and Optalysys, with university work at Georgia Tech, Carnegie Mellon and UC Santa Barbara, and four peer-reviewed papers already built on the toolchain.

Why it mattersUntil now, encrypted inference has been something a handful of cryptographers could build and nobody else could. That is why the privacy answer offered to hospitals, banks and schools has almost always been a promise about policy — we will not look at your data — rather than a system that makes looking impossible. A compiler does not make homomorphic encryption fast: it remains far slower than plain computation, and single-threaded CPU demos are a long way from serving millions of users. What it changes is who can attempt it at all. The question worth watching over the next year is whether any product you actually use ships with this, or whether it stays in the demos.

#Coding

✓ Verified · 3 sources

Apple Music will tell you when a song was made by a machine — but the uploader decides whether to say so
2026-08-21 Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21 Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21 Stripe has just paid $7.5 billion for a model router. Days later Ramp built one and is giving it away until January.
2026-08-21 Meta's assistant is now a Mac app that reads your screen and types into any window — and what it sees can train the model
2026-08-21

Four working demos of AI running on data it cannot read — and a compiler Google wants ordinary developers to use

Related stories