Google details security measures for Chrome's agentic features

December 8, 2025

TL;DR

Browsers are introducing agentic features that can take actions on behalf of users, such as booking tickets or shopping.
These features carry security risks, including potential data or financial loss.
Google is implementing security measures in Chrome, including observer models and user consent.
A User Alignment Critic model, built with Gemini, scrutinizes AI-generated action items.
Agent Origin Sets restrict AI access to specific read-only and read-writeable website origins.
The browser can prevent agents from navigating to harmful or untrustworthy sites.
Users will be prompted for consent for sensitive tasks, such as accessing banking or medical data, or making purchases.
A prompt-injection classifier is used to prevent unwanted actions.
Perplexity has released an open-source content detection model to combat prompt injection attacks.

Continue reading
the original article