What does Gemini think about your brand?

Discovery

Chrome Dev includes a quantized Gemini model for tasks like scam prevention. This analysis examines its on-device execution and reverse-engineered prompts.

Listen

Deep inside the development version of Google Chrome, there is a hidden, local version of Google’s Gemini artificial intelligence model. It is stored directly on your computer as a three-gigabyte file, and it helps with everything from writing assistance to language translation. But one of its most important jobs is stopping online scams.

By looking at the configuration files on the machine, we can see exactly how this scam detection works. Chrome extracts text from the webpage you are visiting and feeds it directly to the local AI. The model acts as a scanner, tasked with answering two key questions: what brand does the page represent, and what is the page trying to do?

The AI is instructed to summarize the page's intent in just one sentence, while being careful not to leak any personally identifiable information. It returns this analysis in a highly structured format. Once the model delivers the brand and the intent, Chrome's built-in classifier takes over to make the final decision on whether the page is trustworthy or a scam. It is a powerful example of how browser security is moving directly onto our devices.

Inside Chrome Dev, there’s a quantized version of Google’s flagship model Gemini for those who have it enabled. The model does many things from summarization, translation, writing assistance all the way to scam prevention. The model definition is a secret, but its weights are stored as a 3GB .bin file on the user machine.

Inside \User Data\optimization_guide_model_store\55\ folder is a file called on_device_model_execution_config.pb which defines a prompt for Gemini’s role in scam detection.

Reverse engineered it looks approximately as this:

-optimization_guide.proto.ScamDetectionRequest
You are a web page text scanner. Your task is to carefully review text from a web page.

The following text is extracted from a web page.
Answer the following questions:
1) What brand does the page represent?
2) Summarize the intent of the page in one sentence. Do not leak PII data.

You should output your answers strictly in the following JSON format, but do NOT use markdown:
{"brand": "<brand>", "intent": "<intent>"}

.optimization_guide.proto.ScamDetectionResponse

The model receives clean text from Chrome and returns two items:

Brand
Intent

Here’s an example of the above implemented with trafilatura and Gemma, a distilled version of Gemini with approximately equal capability as Gemini Nano.

Google’s on-device scam detection classifier then takes over and makes a decision on whether the page is trustworthy or not.

Dan Petrovic · Jan 29, 14:39