What tools detect hidden text and prompt injection in documents before an AI reads them?
Three kinds: free manual checks you already have, tools you paste text into, and tools that read the file itself and compare what each page shows with the text an AI reads.
Of the automatic tools, only the last kind can catch white text, tiny fonts and covered text, because colour, size and position belong to the file, and copying the text throws them away.
The free way to check
The free checks:
- Copy the text out and compare it with the page — see the PDF guide and the Word guide.
pdftotext, a free command-line tool, prints all the text in a PDF.- In Word, set all text to black at a readable size and turn on hidden text.
These cost nothing and find the simple cases. They get slow on long documents, and nothing marks which lines a reader could not see.
Where the free way falls short
- Tools you paste text into find invisible characters and lookalike letters, because those survive a copy. They cannot see white text, tiny fonts or covered text: the paste no longer knows what colour or size anything was.
- Prompt-injection filters read text and judge whether it sounds like an instruction to a model. They do not know which parts were visible, and most hidden prompts do not sound like instructions: in a study of 196,682 real résumés, more than 90% contained no explicit instruction.
- Tools that read the file open the PDF, Word file or web page itself and compare what a reader sees with what the file contains. They range from research projects, such as PhantomLint, to services built to sit in front of an AI, such as HiddenContent.ai.
How HiddenContent.ai checks it
HiddenContent.ai reads the file itself — Word, Excel, PowerPoint, PDF, RTF, web pages, email and plain text — and works out what a reader would see on each page. It reports each hidden passage with where it sits and the trick behind it, and marks hidden text written as an instruction to a model.
When you compare tools, ask each one these questions. Does it check the file, or only pasted text? Does it show you the hidden text and where it sits, or only a score? Does it read the formats you receive, including scanned pages? Does it tell you when it could not read part of a file? Does it send your document to another AI model? HiddenContent.ai checks the file, shows the text and its place, reads scans with OCR, returns “incomplete” rather than “clean” when it cannot read part of a file, and sends nothing to any AI model.
No file leaves our machines, and no AI model reads your document. See one real analysis, end to end for exactly what comes back. Sign up with GitHub for 20 free pages, then a cent a page.
Real cases
- PhantomLint: Principled Detection of Hidden LLM Prompts in Structured Documents — arXiv, 25 Aug 2025
- Measuring Real-World Prompt Injection Attacks in LLM-based Resume Screening — USENIX Security 2026, 27 May 2026
The techniques, one page each
Questions
Can I paste text into a checker to find hidden content?
Only some of it. A paste keeps invisible characters, so a text checker can find those. It loses colour, size and position, so white text, tiny fonts and covered text look like ordinary text.
Do prompt-injection filters catch hidden text?
Not reliably. They judge the wording, not whether anyone could see it, and in a study of 196,682 real résumés more than 90% of hidden prompts contained no explicit instruction.
Which file types does HiddenContent.ai check?
PDF, Word (.docx and .doc), Excel (.xlsx and .xls), PowerPoint (.pptx and .ppt), RTF, HTML, email (.eml) and plain text.
Can I check documents automatically in an AI ingestion pipeline?
Yes. Call the HiddenContent.ai API from your ingestion pipeline before a document reaches the model. It returns a verdict, each hidden passage with where it sits, and clean text you can pass on instead of the original.