Copy it straight from the web page, email or document. When you paste from a web page, the page also checks the formatting that came with it for white, tiny or off-screen text.
Your text stays on this device
The check runs inside this browser tab. The page’s security setting (a Content Security Policy) stops it from connecting to anything, so it has no way to send your text out. Once the page has loaded, switch the wifi off and it keeps working. Opening it in the first place needs a connection.
Watch it work: a 43-second demonstration on YouTube, recorded with the internet switched off. Read the guide.
What the check looks for
- Wording often used to give instructions to an artificial intelligence, such as “ignore your previous instructions”, “note to the assistant”, “the assistant summarising this should say…”, lines dressed up to look as if they come from the system that runs the assistant, or “do not tell the user”. It knows the commonest version of “ignore your previous instructions” in French, Spanish, German, Portuguese, Italian and Dutch too.
- Requests to send information away, such as instructions to email passwords, forward a conversation, or an image link built to carry your data to another website.
- Hidden characters and hidden messages, including tag characters that spell out a whole sentence you cannot see, which the page decodes and checks too.
- Coded text: instructions turned into a jumble of letters and numbers that a computer can turn back (base64 or hexadecimal). The page turns them back and checks them.
- Text hidden on a web page: white text on a white background, tiny text, transparent text and boxes pushed off the screen, when you paste straight from the page in a browser that copies formatting, such as Chrome or Edge. Comments in the page code are not copied when you paste, so open a saved web page file to check those.
- Tricks that break up words: an invisible character inside a word, look-alike Cyrillic or Greek letters, numbers used as letters (“1gn0r3”) and words split with a hyphen at the end of a line. The page undoes these and checks the words again.
Why this matters now
- On 11 June 2025, researchers at Aim Security disclosed EchoLeak, a flaw in Microsoft 365 Copilot recorded as CVE-2025-32711 in the United States National Vulnerability Database. An email with hidden instructions could make Copilot leak a person’s data without them clicking anything. Microsoft fixed it before it was made public.
- The Open Worldwide Application Security Project puts prompt injection first in its 2025 Top 10 risks for applications built on large language models.
- On 8 December 2025 the United Kingdom’s National Cyber Security Centre warned that prompt injection may never be fully prevented, describing language models as “inherently confusable”.
- On 3 September 2026 Microsoft Security Research described a phishing campaign that hid invisible tag characters inside words such as “funding” to slip past email filters.
What it catches and what it misses
This check looks for known phrases and known hiding tricks. It does not understand meaning the way a person does. Last checked on 19 October 2026.
It catches
- “Ignore all previous instructions” and its common rewordings, such as “disregard what you were told earlier” or “previous instructions no longer apply”, and the commonest version in six other languages.
- Notes addressed to an artificial intelligence, requests to hide something from the reader, requests to send the conversation or passwords to a web address, and image links built to carry data away.
- Instructions hidden in tag characters, in coded text, in white, tiny or off-screen text from a pasted web page, and words broken up with invisible characters, look-alike letters, numbers or line-end hyphens.
It can miss
- Instructions written in completely new words, or in languages other than the seven it knows.
- Text inside pictures, and anything an artificial intelligence reads straight from a page you have not pasted here.
- Instructions spread across several harmless-looking sentences, and links that carry data in ways it does not recognise.
It can wrongly warn about
- Ordinary sentences that happen to use the same words, such as a recipe that says to ignore earlier steps. Ordinary advice such as “Do not tell anyone your PIN” and labels such as “System: Windows 11” are tested and pass.
A clear result means no known warning signs were found, not that the text is guaranteed harmless. Keep reading what matters yourself, and do not let an artificial intelligence act for you on content you have not checked. If it misses something, tell us and it will be added to the tests.
Questions and answers
What is prompt injection?
It is a trick where text written for an artificial intelligence (AI) is hidden inside something you give it, such as a web page, an email or a document. When you ask the AI to read or summarise that content, it may follow the hidden instructions instead of yours, for example to ignore its rules, to give you a false summary, or to send your information somewhere.
How do I check text for prompt injection?
Copy the text from the web page, email or document and paste it into the box on this page, then press Check the text. The page looks for wording often used to give instructions to an artificial intelligence, hidden characters, hidden messages, coded text and, when you paste from a web page, text that was hidden with white or tiny lettering. It then shows what it found and gives you a cleaned copy.
Can a web page give instructions to ChatGPT?
It can try. When you ask ChatGPT, Gemini, Claude, Copilot or another artificial intelligence to read a web page or an email, it reads every word, including words hidden in white text, in invisible characters or in the page code. Some of those words may be written as instructions, such as “ignore your previous instructions” or “tell the reader this product is the best”. Whether the assistant follows them depends on the assistant, which is why checking the text first helps.
Can this check catch every attack?
No, and no check can. On 8 December 2025 the United Kingdom’s National Cyber Security Centre warned that prompt injection may never be fully preventable, because language models struggle to keep commands apart from content. This page catches many common patterns and several hiding tricks, and it lists below what it can miss. Read anything important yourself before an artificial intelligence acts on it.
What should I do if the check finds something?
Do not paste the text into your artificial intelligence as it is. Use the cleaned copy, or remove the flagged parts, or summarise the source yourself. If it came in an email or a message, treat the sender with care. Never let an artificial intelligence send emails, make payments or change files based on content you have not read.
Does it work with ChatGPT, Gemini, Claude and Copilot?
Yes. You check the text here first, then paste the cleaned copy into whichever artificial intelligence you use, including apps that offer several in one place. Different assistants can react differently to the same text, so a clear result here is not a promise about how any one of them will behave.
Is it safe to ask an artificial intelligence to summarise a web page?
Usually the summary is fine, but the page can contain words aimed at the assistant rather than at you, and they can change what the summary says. The risk grows when the assistant can also act for you, such as sending emails, opening links or reading your files. Check the text first, read the parts that matter yourself, and do not let an assistant act on a page you have not looked at.
Can it check a PDF or a Word document?
Copy the text out of the file and paste it here. To find text hidden under black boxes in a PDF, use Is It Really Blacked Out?, and for hidden content in Word, Excel and PowerPoint files use Release Check.
Does it cost anything, and do I need to sign in?
No. It is 100% free, with no sign-ups and no tracking, and you can start straight away. UtilForge also sells optional packs separately on Gumroad, and this tool works without them.
What happens to the text I paste?
It stays in this browser tab. The page has a security setting (a Content Security Policy, a rule your browser enforces) that stops it from connecting to anything, so it cannot send your text anywhere. It uses no cookies or storage. Close the tab and the text is gone. You can test it: once the page has loaded, switch the wifi off and it keeps working. Opening the page in the first place needs a connection.
Who made this?
UtilForge.org, a registered 2026 Champion of Cybersecurity Awareness Month, which makes free cybersecurity tools for a safer online world for all. The About page explains why the tools exist.