Using AI to stop tech support scams in Chrome (security.googleblog.com)

Chrome 137 adds an on-device defense against tech support scams by running the Gemini Nano large language model locally to generate signals for Google Safe Browsing. When heuristic triggers typical of support scams (for example, use of the keyboard lock API, full-screen takeovers, or blocked input) are detected, Chrome feeds the page content to Gemini Nano to extract intent and other security signals; those summarized signals are then sent to Safe Browsing for a final verdict and, if warranted, an interstitial warning. The approach aims to catch short-lived, evasive pages that often disappear within minutes and that may render differently for real users than for crawlers. Technically, Chrome limits LLM use to preserve performance and privacy: the model runs asynchronously on-device with token limits, throttling, GPU quotas, and sparing triggers. LLM-derived signals are only transmitted for users who opt into Enhanced Protection, while Standard Protection users gain indirectly as newly discovered threats are added to blocklists. Google plans to extend the same pipeline to other scam types (package-tracking, unpaid toll fraud), roll it out to Android, and research defenses against prompt-injection and timing-bypass attacks. For the AI/ML community, this is a concrete example of on-device LLMs improving real-time security telemetry and combating crawler-evasive, ephemeral threats.