What changed in the Hidden Text Detector

The earlier version checked pasted text for a narrower set of unusual characters. The upgraded Hidden Text Detector now accepts file uploads and checks for a wider range of signals, including concealed formatting and suspicious instructions aimed at AI systems.

You can still paste text directly into the tool. For documents, the checks vary by file type.

What you can scan

Upload a PDF, DOCX, TXT, or Markdown file, or paste text into the detector.

For PDFs, the tool checks text spans for signs of concealment, such as unusually small, low-contrast, transparent, or off-page text. It also looks for patterns that resemble instructions aimed at an AI system.

For DOCX files, checks include hidden or “vanish” formatting, text containers, and style or layout clues. Microsoft’s Word documentation describes how text can be marked as hidden through formatting.

Pasted text and TXT or Markdown files are checked for unusual or invisible Unicode characters and mostly English prompt-injection signals. Prompt injection can appear in content an AI system is asked to read, as well as in a direct request. See OpenAI’s overview of prompt injection and its agent-builder safety guide.

Review findings before acting

Findings include severity, evidence, and location to help you inspect the relevant content. You can review or ignore a finding. If the tool offers to remove something, removal is your choice; it never happens automatically. Some content cannot be safely removed. Cleaned output is separate from the source file, which remains unchanged. Review the result before using it.

Know the limits

The PDF scan checks text spans. It does not use OCR to read text embedded in scanned pages or images, and its PDF checks do not scan Unicode controls. Detection is not guaranteed, and the tool is not an antivirus or malware scanner.

Some unusual characters and formatting are legitimate. Bidirectional controls, for example, can help display right-to-left text correctly; the Unicode Bidirectional Algorithm FAQ explains how text direction is handled. Arabic and other scripts, emoji, typography, and quoted discussions of suspicious instructions can produce findings that need context.

Scans run on the app’s servers, including scans of pasted text.

Try the upgraded detector

Open the Hidden Text Detector to upload a supported file or scan pasted text, then review the findings before sharing or using the content.