← Blog

Security · Field note 037

Google Messages sends complicated scams to a private AI cloud

An optional protection mode combines cloud Gemini analysis, on-device link checks, and clearer warnings for unknown senders.

A smartphone and glowing shield blocking red phishing messages

Google is giving Messages a stronger defense for scams that look convincing one message at a time.

An optional enhanced protection mode will use Gemini models in Google’s Private AI Compute environment to evaluate the wider context of conversations with unknown senders. The goal is to recognize a scammer’s intent earlier, including unfamiliar attacks designed to evade the rules and on-device models already in the app.

Google says Private AI Compute encrypts the data and processes it inside a protected environment that the company cannot reuse or access for another purpose. The cloud analysis supplements existing on-device scam detection rather than replacing it.

Links get checked before the tap

Enhanced protection also uses Gemini Nano on the phone to evaluate unfamiliar links against real-time web signals. The feature is initially available on newer devices including the Pixel 10 family, Samsung’s Galaxy S26 series, Galaxy Z Flip 8, Galaxy Z Fold 8, and other selected phones.

Warnings for links from numbers outside a user’s contacts are being redesigned as well. Google says the prompts will provide clearer context and create a more deliberate pause before someone opens a suspicious destination. The enhanced mode is opt-in and will roll out over the coming weeks.

The protection arrives as scams increasingly unfold through conversation rather than a single obviously malicious message. A fake recruiter, bank representative, or delivery agent can spend several exchanges establishing trust before introducing a link or requesting an action. Looking at the sequence gives a detection system more to work with than scanning each line in isolation.

Android’s defenses are widening

Google is also expanding on-device scam detection for chat notifications beyond Pixel to phones from Honor, Motorola, OnePlus, Oppo, Xiaomi, Vivo, and several other manufacturers.

Separate Android safeguards will hide sensitive information such as passwords and banking details during screen sharing in Chrome and WebView on Android 15 or newer. Beginning in early December, Pixel and selected Android phones will also stop apps from quietly turning on call forwarding — a technique that can redirect calls containing one-time passcodes. Android will warn users if someone persuades them to type a forwarding code manually.

The design tradeoff is familiar: spotting more sophisticated fraud can require more context, while private conversations deserve strict limits on where that context goes. By making the stronger analysis optional and routing it through a locked-down compute system, Google is trying to improve detection without treating ordinary messages as freely available cloud data.