We're building an open-source safety layer for AI browser agents (24 stars in, here's what we've learned)

A few weeks ago my team at PixieBrix shipped something that might become essential with the future of agentic web browsing: Agent Browser Shield, an open-source layer that sits between a browser-use AI agent and the web.

Here's the problem that started it. When an AI agent browses a page, it reads everything. The cookie banner, the ads, the fake "only 2 left!" scarcity, and the prompt injection someone hid in an HTML comment. It can't tell the task from the junk. That's wasted tokens at best and a security hole at worst.

So the Shield cleans the page before the agent sees it. It masks PII and secrets, strips prompt-injection surfaces, and pulls out dark patterns and noise. Fewer tokens burned, less chance of getting manipulated.

We're 24 GitHub stars in, packaged for Browserbase, and out as an OpenClaw skill. Still very much building in the open, the roadmap isn't locked and I'd genuinely rather shape it with people who run agents against real sites.

If you've deployed a browser agent, what's the gnarliest thing a page has done to it? That's the stuff we want to defend against next.