Architecting Security for Agentic Capabilities in Chrome
- Source:
- Google Online Security Blog
- Published:
- Fetched:
- Updated:
- Affects:
http://security.googleblog.com/2025/12/architecting-security-for-agentic.html ↗
Posted by Nathan Parker, Chrome security team Chrome has been advancing the web’s security for well over 15 years, and we’re committed to meeting new challenges and opportunities with AI. Billions of people trust Chrome to keep them safe by default, and this is a responsibility we take seriously. Following the recent launch of Gemini in Chrome and the preview of agentic capabilities , we want to share our approach and some new innovations to improve the safety of agentic browsing. The primary new threat facing all agentic browsers is indirect prompt injection . It can appear in malicious sites, third-party content in iframes, or from user-generated content like user reviews, and can cause the agent to take unwanted actions such as initiating financial transactions or exfiltrating sensitive data. Given this open challenge, we are investing in a layered defense that includes both deterministic and probabilistic defenses to make it difficult and costly for attackers to cause harm. Designing safe agentic browsing for Chrome has involved deep collaboration of security experts across Google. We built on Gemini's existing protections and agent security principles and have implemented several new layers for Chrome. We’re introducing a user alignment critic where the agent’s actions are vetted by a separate model that is isolated from untrusted content. We’re also extending Chrome’s origin-isolation capabilities to constrain what origins the agent can interact with, to just those that are relevant to the task. Our layered defense also includes user confirmations for critical steps, real-time detection of threats , and red-teaming and response . We’ll step through these layers below. Checking agent outputs with User Alignment Critic The main planning model for Gemini uses page content shared in Chrome to decide what action to take next. Exposure to untrusted web content means it is inherently vulnerable to indirect prompt injection. We use techniques like spotlighting that direct the model to strongly prefer following user and system instructions over what’s on the page, and we’ve upstreamed known attacks to train the Gemini model to avoid falling for them. To further bolster model alignment beyond spotlighting, we’re introducing the User Alignment Critic — a separate model built with Gemini that acts as a high-trust system component. This architecture is inspired partially by the dual-LLM pattern as well as CaMeL research from Google DeepMind. A flow chart that depicts the User Alignment Critic: a trusted component that vets each action before it reaches the browser. The User Alignment Critic runs after the planning is complete to double-check each proposed action. Its primary focus is task alignment: determining whether the proposed action serves the user’s stated goal. If the action is misaligned, the Alignment Critic will veto it. This component is architected to see only metadata about the proposed action and not any unfiltered untrustworthy web content, thus ensuring it cannot be poisoned directly from the web. It has less context, but it also has a simpler job — just approve or reject an action. This is a powerful, extra layer of defense against both goal-hijacking and data exfiltration within the action step. When an action is rejected, the Critic provides feedback to the planning model to re-formulate its plan, and the planner can return control to the user if there are repeated failures. Enforcing stronger security bound aries with Origin Sets Site Isolation and the same-origin policy are fundamental boundaries in Chrome’s security model and we’re carrying forward these concepts into the agentic world. By their nature, agents must operate across websites (e.g. collecting ingredients on one site and filling a shopping cart on another). But if an unrestricted agent is compromised and can interact with arbitrary sites, it can create what is effectively a Site Isolation bypass. That can have a severe impact when the a…