Prompt Injection Chrome Extension: Inspect Hidden Page Text

Before you ask an assistant to summarize a page, inspect suspicious hidden text and check what your installed extension transmits. The Web Store listing and checked-in source describe different versions; the setup checklist separates those behaviors.

Use the extension findings as prompts for review. A hidden instruction can influence a summary or an action if the assistant’s extraction process includes it; a warning doesn’t control that independent retrieval.

Review the installed version before scanning a page

1. Install from the Chrome Web Store link above and review its requested site access.
2. Check the installed version and its settings. If it uses the API behavior in the version 2 source, get a free SafePrompt key and configure it in the popup.
3. Review the highlighted text and extension findings before asking an assistant to use the page.

The checked-in version 2 manifest requests storage, activeTab and access to all URLs. Its content script sends suspicious text to the SafePrompt API when a key is configured. Read the privacy policy before using it on sensitive pages, and limit site access through Chrome’s extension settings.

Does a browser warning block an AI assistant?

A warning does not control ChatGPT, Claude or another assistant’s independent retrieval. The extension does not prove that every hidden element was found or that a page without warnings is safe. App builders should screen the exact extracted content their model receives and restrict the actions that content can influence.

The Chrome Web Store listing and the checked-in implementation describe data transmission differently. This article discloses the API check present in the source; verify the installed version’s behavior before using it with confidential content.

What it protects you from

Hidden text injection hides instructions on a web page that a person may not see. An assistant is exposed when its extraction process includes those instructions. When you ask an assistant to summarize or act on that page, the hidden instructions can hijack the AI's response without you ever knowing they were there.

How it works

The extension scans each page you open for the CSS tricks that hide text from you but not from an AI: colour-matched text, white-on-white, display: none, visibility: hidden, and zero-opacity elements. The wider catalogue of techniques is in our complete guide to hidden text injection. The checked-in extension highlights candidate elements. Hidden styling also supports menus, screen-reader text and other ordinary UI, so inspect what the text instructs the assistant to do before calling it an attack.

The checked-in version 2 source runs a pattern pass in your browser and sends suspicious text to the SafePrompt API when a key is configured. The Web Store listing for version 1.0.0, updated February 11, describes local processing without external transmission. These are different version descriptions; the checked-in source does not establish what the store-installed extension transmits.

Who this is for

• Anyone who summarizes pages with AI. If you routinely ask ChatGPT, Claude, Gemini, or Perplexity to read a page for you, this is the surface you are exposed on.
• Security-curious browsers. Inspect hidden text on real pages and distinguish attack instructions from ordinary interface content.
• Builders testing their own pages. Check whether your site accidentally carries hidden text an attacker could exploit against your visitors' AI.

The extension and the API cover two different surfaces

The extension gives you, the person browsing, a warning to review. If you are building an app where users send prompts to your model, that is a different door, and it needs the SafePrompt API: one call to validate a prompt before it reaches your model, benchmarked publicly against real attacks. Free plan, no card, $29/mo when you outgrow it. Same threat, two surfaces, two tools.