AgentGlass preflight

// about

About AgentGlass

AgentGlass is a free scanner that checks a public web page for hidden instructions aimed at AI: the kind of prompt injection that can hijack an AI agent reading the page on your behalf. It started with one specific problem, people using AI to apply for jobs.

Why it exists

An AI agent reads a page differently from a person. It takes in everything: the text you see, the text hidden by CSS, HTML comments, metadata and the structured data written for search engines. A job posting can use any of those places to plant instructions for the agent, and the person who asked the agent to apply never sees them.

Nobody was checking the page before the agent acted on it. AgentGlass does that one check.

How it works

You paste a public URL. AgentGlass loads it in an isolated headless browser, the same way an agent would, and looks for content that is hidden from a human reader or phrased like an instruction: hidden text, HTML comments, metadata and script or JSON-LD blocks. It returns a risk level, a short verdict and the exact evidence it found.

What it is not

AgentGlass identifies patterns, not intent. Hidden elements can be perfectly legitimate: responsive layouts, accessibility text and templates all hide things for good reasons. A high-risk result is a signal that a person should read the page before an AI submits anything, not proof of fraud. It is also not a general job-scam detector; it focuses on the specific problem of instructions written for AI.

Principles

Who builds it

AgentGlass is built and run by Puchidex. It is currently in public beta.

Contact

Questions, feedback or a page you think we should look at: hello@puchidex.com. For security issues, please read the security page first.