Google is adding an optional Google Messages mode that sends context from conversations with unknown senders to Gemini models in the cloud. Announced on October 14 and due to roll out over the coming weeks, Enhanced protection is designed to spot slow-building scams before a request for money or private information makes the danger obvious.
The bargain is clear, but the data boundary is not. Turning on Enhanced allows part of a private exchange to leave the phone so a larger model can examine it. Google says Private AI Compute is designed to keep that data inaccessible to anyone else, including Google. The company has not specified how much of the conversation is sent.
Enhanced sends an unspecified payload representing conversation context to cloud Gemini. Private AI Compute encrypts the transfer and uses remote attestation, which checks that the receiving system is running approved software. The data is decrypted inside protected computing environments for analysis, and a response returns to the phone. In a Google-commissioned assessment of Private AI Compute, NCC Group spent 100 person-days reviewing the architecture and selected code and configurations. It found strong barriers against outsiders and malicious insiders, but excluded phone apps and the underlying confidential-computing platform. The review therefore supports important parts of the cloud design, not the Google Messages integration from end to end.
NCC also qualified the promise that even Google cannot access the data. It said Google as an organization ultimately retains the power to change its centralized system in ways that could expose information, although the reviewers found no evidence that Google had done so. Some verification also depended on demonstrations and internal access supplied by Google.
Enhanced is separate from ordinary spam protection, which is enabled automatically. Google described its 2025 conversational scam detector for non-contacts as operating on the phone. But turning Enhanced off does not make every Messages check local: Google's current documentation says Messages may upload URLs for safety checks and briefly process unencrypted message content on servers when a device cannot filter it locally. Those are different data flows from the new contextual Gemini analysis.
Choosing to report spam is another separate action that sends message content to Google. Google's public pages do not say how many messages Enhanced examines, whether attachments or metadata are included, how long information is kept, whether it is used for training, what remains after a user opts out, or whether the other participant is notified. Those omissions do not prove that such controls are absent.
Enhanced itself bundles several checks, while Google's wider announcement adds protections with different schedules and hardware requirements. A phone that qualifies for cloud scam detection will not necessarily qualify for the on-device Gemini Nano link analysis.
| Protection | Timing and consent | Published device requirements |
|---|---|---|
| Cloud Gemini scam detection in Messages | Rolling out over the coming weeks; requires an Enhanced protection opt-in. | Pixel 6 and newer; Galaxy S24 series and newer; Galaxy Z Flip6 and Z Fold6 and newer; other phones with Android's AICore AI service. |
| On-device Gemini Nano checks for unfamiliar links | Part of optional Enhanced protection. | Pixel 10 and newer; Galaxy S26 series; Galaxy Z Flip8 and Z Fold8; other select devices. |
| Extra warnings before opening unfamiliar links | Part of optional Enhanced protection; previously limited to selected countries or Android Advanced Protection. | Android 6 and newer. |
| Scam alerts in chat notifications | Expansion beyond Pixel later in 2026; the announcement does not specify additional consent steps. | Participating manufacturers; models vary. |
| Screen-share privacy in Chrome and in-app pages using Android WebView | Rolling out in October; automatically hides login fields and payment screens during sharing. | Android 15 and newer. |
| Call-forwarding safeguards | Starting in early December; blocks silent background activation and warns before manually entered forwarding codes take effect. | Pixel and other select devices. |
Google has not published a country list for Enhanced protection or a complete list of the AICore and other select devices. The table describes announced eligibility, not confirmation that every qualifying phone already has each feature.
Google's bank-call defense uses a different mechanism: verification by the bank rather than a Gemini judgment. Verified financial calls requires Android 11 or newer, a participating bank's official app and a previous sign-in. For supported calls, Android asks the app whether the bank initiated the call and ends the call if the bank cannot confirm it. Google says the system has blocked more than 3 million calls spoofing inbound-only numbers belonging to Banco do Brasil and Itaú Unibanco since June. Not every bank or call is supported.
Google also says its existing protections flag more than 5 million suspected scam calls and texts each month. That shows activity, but without totals for missed scams and mistaken warnings, it does not establish accuracy or money saved. A Google-funded Leviathan comparison tested four phones for a month on T-Mobile using free, built-in features enabled by default. Its full report was dated one day before Enhanced was announced and excluded penetration testing, so it cannot validate the optional cloud detector.
Enhanced protection could be valuable if it catches a slow-building con before money or credentials change hands. Google has not yet published the comparison users need to assess that tradeoff: how much more the cloud mode catches than standard protection, how often it warns about legitimate conversations, and precisely what context users must send to gain that protection.