This post is also available in:
עברית (Hebrew)
Google is rolling out new anti-scam capabilities in its Chrome browser, introducing a lightweight on-device AI model designed to spot fraudulent websites and alert users in real time. The technology, powered by Google’s Gemini Nano large language model (LLM), will debut in Chrome 137, which is currently in beta and expected to launch publicly this week.
The update represents a significant step in browser-level threat detection, particularly against fast-evolving scams. These types of malicious websites often exist for less than 10 minutes – too brief for traditional web crawlers to identify. By operating locally on users’ devices, Gemini Nano enables immediate analysis at the moment a user accesses a site.
The system works by scanning for features commonly associated with scams, including deceptive language and design patterns. It then generates security signals, which are relayed to Google’s Safe Browsing service to determine whether the site should be flagged. If identified as dangerous, Chrome will display a full-screen warning to the user.
Because it runs on-device, Gemini Nano doesn’t compromise user privacy. Google notes that the feature is only available to users who opt into the Enhanced Protection setting within Chrome’s Safe Browsing controls. Detected scam sites will also be added to broader blocklists, indirectly benefiting users on the Standard Protection tier.
In addition to desktop protection, Google plans to extend the AI-powered system to Chrome for Android later this year. Another related feature coming to Android will use local machine learning to monitor notification content, flagging messages from websites that may be abusive or deceptive. Users will receive alerts and be able to unsubscribe from such notifications with a single tap.
These enhancements build on Google’s broader push to integrate AI-based security across its ecosystem. Similar tools have already been deployed in Google Messages and Phone apps to detect scam calls and texts before they reach the user.
