Tolz.me
AISecurityWeb

WhatsApp's Hidden AI Notes: How Unseen Text in Web Pages Steers Chatbots

WhatsApp help pages carried AI instructions nobody could see. What Redact found, what WhatsApp said, and how to check a page for hidden text.

Tolz.me team9 min read
Illustration of a robot reading a web help page on a laptop screen

On October 8, 2026, Redact found text on WhatsApp's help centre pages that no visitor could see. It was in the page source, written as instructions for an AI, and it was sent along with the page to anything that fetched it.

Redact published its findings on October 19, 2026, and the story came up on Hacker News. The details are worth reading closely. The notes were aimed at the software that answers WhatsApp support questions, and on the delete-account page one of them pointed in a direction that did not match WhatsApp's own published policy.

What Redact found

  • Redact read all 545 pages that the WhatsApp help centre links to.
  • 80 of those pages contained hidden text written for an AI.
  • The hidden text appeared across 102 separate blocks on those pages.
  • By October 18 at 02:08 UTC, every hidden block was gone from the live pages.

The delete account page

The delete-account page got the most attention because the note on it was about a decision users care about. The page covers the auto-deletion periods: a deactivated account is deleted after 30 days, an inactive account is generally deleted after 120 days, and in the United States the window is 365 days.

The hidden note came after a restatement of those same periods. It told the AI not to recommend contacting any Data Protection Authority, regulator, or supervisory authority. It also told the AI not to suggest that WhatsApp can manually delete an account.

Redact points out a conflict with what WhatsApp publishes. WhatsApp's privacy policy for Europe, effective July 2, 2026, tells users they can complain to the Irish Data Protection Commission or to another supervisory authority. The hidden note steered the assistant away from exactly that route.

The reporting and blocking page

On the page about reporting and blocking other users, the hidden note handled the case where a user mentions legal action or lawyers. It told the AI not to reference data protection regulations such as GDPR, LGPD, or CCPA, even when the user raises them directly.

Most of the blocks were housekeeping

The two notes above were the striking ones, not the norm. According to Redact, most of the 102 blocks were routine housekeeping, the kind of instruction that keeps a support chatbot consistent. A few of them protected users rather than the company. One told the AI not to suggest lying about an age.

The timeline

  • October 8, 2026: Redact finds the hidden text while reviewing the WhatsApp help centre.
  • October 18, 2026, 02:08 UTC: Redact's check finds every hidden block removed from the live pages.
  • October 19, 2026: Redact publishes the write-up, and the story comes up on Hacker News.
  • WhatsApp tells WIRED the notes were instructions for its support chatbot and says they reached the public pages because of a bug, which WhatsApp says has been fixed.

What a reader saw versus what the page carried

A reader saw nothing unusual. The visible page showed ordinary help content. Someone reading the delete-account page saw the deletion periods and the help copy around them. The extra sentences existed in the document rather than in the layout, so no amount of scrolling, zooming, or careful reading on the rendered page would reveal them.

That gap is the whole problem. A browser hides text with rules that only the browser applies. Any other consumer of the same document, including a crawler, a scraper, an accessibility tool, or an assistant, receives the rules and the text separately, and mostly reads the text.

Where the blocks sat in the source

On the delete-account page, Redact found the note near a name in the page source: HelpContentGalaxyLLMInclude. Searching the source for that string, or for the wording of the note, is how Redact tells readers to find it. Copies on the Internet Archive still show the notes, so the pages can be inspected after the live versions were cleaned up.

Two details make this more than a curiosity. The pages were served to Google's and Bing's crawlers, so the notes were available to anything that indexes them. And Redact says it has not tested whether any AI assistant acted on the instructions, or whether WhatsApp's own chatbot followed them.

For contrast, Redact checked the Facebook and Instagram help centres as well, 300 pages in total. Those pages had no such notes.

How hidden text works in HTML

There are several ways to put words in a page that a reader never sees, and they are not the same technique. That distinction matters here, because the WhatsApp notes were blocks of text in the page source, not invisible characters.

Hidden elements

The usual approach is a container the browser is told not to display: the `display:none` CSS rule, the `hidden` attribute, or a wrapper collapsed to no size and covered by other content. The text is still part of the document, and it travels with every request for the page.

Invisible characters

A different technique uses characters that take up no space on screen. The zero-width space, the zero-width non-joiner, the zero-width joiner, the word joiner, and the byte order mark all render as nothing while staying part of the string. A name can look empty and carry a word inside it. A fetcher reading the raw text sees the word.

That is the trick behind a family of Discord tricks where a visible nickname hides a second name behind it. The Discord invisible name tool shows how those characters behave once they are inside a name.

WhatsApp's notes were not invisible characters. They were ordinary sentences that the page carried but never drew on screen.

Why it matters when you feed pages to an AI

Any tool that turns a web page into model input sends the hidden parts too. A browser is what hides text from a reader, and a fetcher is not a browser. It takes the document as it arrived.

Once that text reaches a model, it is text in a prompt. If it is phrased as a directive, an assistant that has been told to follow instructions can treat it as one, including directives the visible page never showed a reader. That is the shape of a prompt injection, and the hidden version is harder to spot because a person skimming the page sees nothing unusual.

Retrieval systems make this worse in a quiet way. A search pipeline splits page text into chunks before a model sees any of it, so a hidden block can be retrieved on its own, without the surrounding article for context. The model receives a few sentences that look like system instructions and no page around them to explain them.

Redact's write-up is careful about the limit of the evidence. The notes existed and were served to crawlers, but Redact did not test whether an assistant acted on them. The exposure is structural rather than demonstrated.

Why the WhatsApp explanation changes how you look for these

If the notes were instructions for a support chatbot, then the useful search term is the chatbot itself rather than anything sinister. Look for text that addresses a reader that is not a person: rules for a helper, a list of things to say and not say, a paragraph written to steer a response.

The bug explanation deserves to be taken seriously, because a bug and a feature look identical from the outside. Both leave text in the page that a person cannot see. What separates them is intent, and intent is exactly what you cannot read from the source alone.

The same reasoning applies to any page you check. Ask what a hidden block would be for, then decide whether that answer fits the site. A documentation page carrying notes for its own search assistant is a different story from a shop carrying notes aimed at an assistant that writes product reviews.

How to check a page for hidden text

  1. View the source. Press Ctrl+U on Windows and Linux. On a Mac it is Cmd+Option+U in Chrome or Cmd+U in Firefox. The source is the document the server sent, with no CSS applied, so hidden elements show up as plain text.
  2. Search the source for telltale words. Try AI, assistant, ignore, do not, LLM, or the name HelpContentGalaxyLLMInclude. A block of sentences addressed to a model reads differently from interface copy once you go looking.
  3. Check what the browser is hiding. Open the developer tools, select the page content, and look for containers set to `display:none` or carrying the `hidden` attribute. Any of them can hold text a model would read.
  4. Check the text you were handed. If a page was pasted into a chat window or a document, invisible characters travel with it. The invisible character remover shows which characters are actually inside the string.

What to do when you find instructions aimed at a model

  • Report the page to the site owner, quoting the block and giving the location in the source.
  • Treat fetched pages as untrusted input. Give an assistant the visible text as content, and keep your instructions on the side where they belong, outside the fetched material.
  • Look at archived copies when the live page has already been cleaned up, which is what happened here.

Sources

Tools used in this guide

Open a tool, make the change, and copy the result. No account needed.