AI agents are one prompt injection away from doing something you'd never ask them to do. We built a fix.

There are 6 billion internet users today. With AI agents, we're headed toward hundreds of billions of "users" hitting the web. And if you're using OpenClaw to browse, your agent has basically zero protection from the stuff humans have learned to watch out for, like fake reviews, dark patterns, hidden instructions baked into page content designed to hijack what it does. My team at PixieBrix has spent years helping BPO contact centers protect against insider risk and social engineering in the browser. So we took that same thinking and applied it to AI agents. **The result is Agent Browser Shield, a free, source-available browser extension we just put on ClawHub and GitHub.** Install it alongside your OpenClaw setup and it rides along with your agent in the sits between the browser and your agent, blocking: * Prompt injection (visible and hidden instructions in page content) * Dark patterns (UI designed to manipulate or coerce) * Context pollution (low-value junk that throws off instruction following) Bonus: stripping irrelevant content also cuts down on tokens used. More secure and cheaper to run! This is not a solved problem though. We're building this in the open because agentic browsing is evolving faster than any one team can track. We're measuring the improvements and working on sharing benchmarks as we go. We want the community to help us figure out where to focus. Come build this with us! 🦞 Star on ClawHub: [https://clawhub.ai/pixiebrix/agent-browser-shield](https://clawhub.ai/pixiebrix/agent-browser-shield) ⭐ Star the GitHub: [https://github.com/pixiebrix/agent-browser-shield](https://github.com/pixiebrix/agent-browser-shield) What weird stuff have you seen your OpenClaw agent do on a page? Curious what you're running into so we can build more rules.