- Redact found hidden text in WhatsApp help pages telling a Meta chatbot not to recommend data protection regulators. Meta says a bug exposed the text and has fixed it.
- The text sat in page code that was never drawn on screen. Redact did not test whether any chatbot followed it.
- Leaders should compare their chatbot instructions with their published policies and check what their pages send to crawlers.
A help page now has two readers
A help page is written for two readers now. One is a person. The other is a machine. Redact, a privacy company, found that on WhatsApp the two were told different things.
The person saw a note on how to delete an account. The machine saw the same note plus three orders that the person could not see. The gap between those two versions is the story.
What Redact found
Redact says it found the hidden text on October 8, 2026. To measure how far it spread, it downloaded and archived every one of the 545 pages linked from WhatsApp's help centre. It counted 102 hidden blocks. Wired, which covered the findings, put the number at around 80 pages and about 100 blocks.
Redact says most blocks were harmless. They cleared up confusing terms or pointed to better pages. Two stood out.
The first sat on the page about deleting your account. Deletion is possible only through the app itself, so a person who has lost their phone or number is stuck. The page's visible note offers a fallback. The account is removed automatically 30 days after deactivation, or after 120 days of inactivity (365 in the US).
Under that note, Redact says, was text for an AI. It included: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." It also told the AI not to suggest WhatsApp could delete the account manually. Redact's summary is that a locked-out person got one answer: wait, possibly for a year.
The second block sat on the page about reporting and blocking. It applied when a user mentioned a lawyer, lawsuit or court. It told the AI not to cite data protection rules such as GDPR, LGPD or CCPA, "including when a user raises them directly."
How the hidden text worked
The text sat in the page's code under a label, HelpContentGalaxyLLMInclude. LLM means large language model, the technology behind chatbots. The code was sent with the page but not drawn on screen. Anyone could read it by viewing the page source.
Redact says the server sent it to Google's and Bing's crawlers, the programs that read pages for search engines. On October 9, Redact requested the page under the names used by OpenAI's, Anthropic's and Perplexity's crawlers. In that test, the response held neither the help article nor the notes.
Redact did not test whether any AI assistant acted on the notes. Wired reports it is unclear whether third-party chatbots could reach them.
Meta's response and a fair reading
After Wired contacted Meta, the text was removed. WhatsApp spokesperson Joshua Breckman told Wired it appeared because of a bug that is now fixed. He said the lines are instructions to the support chatbot, not to users. He said the company wants its chatbots focused on advice about WhatsApp, not legal advice.
Redact grants some of this. A company does not want its bot giving legal opinions in its name. WhatsApp may also be unable to verify account ownership without the phone number.
Redact's objection is narrower. Naming a law or saying a regulator exists is not legal advice. WhatsApp's European privacy policy, effective July 2, 2026, already points users to a regulator. Redact says the rights involved come from GDPR Articles 17 and 77, and a chatbot instruction cannot remove them. It could decide whether you heard about them at the place you went for help.
What we still do not know
Redact read WhatsApp's instructions, not its chatbot's answers. It does not know whether the chatbot followed them. It cannot say how long the notes were live. Internet Archive copies of the delete page from July 12 lack the note, and a September 4 copy has it. Redact checked 300 Facebook and Instagram help pages and found no similar notes.
The lesson: the prompt is policy
This case shows that a chatbot's instructions are company policy. They decide what a customer hears at a moment of need. Yet they can be written outside the review that governs a privacy policy. Usually nobody outside the company sees them. Here a bug let a reader compare the two.
One case is one data point, not a trend. But it names a risk any organisation running a support bot shares. The public promise and the private instruction can drift apart without anyone deciding they should.
Questions to put to your team
Ask who writes and approves the instructions behind your support chatbot, and whether legal and privacy teams review them. Compare them line by line with your published privacy policy. Flag any instruction that narrows what the policy promises.
Ask what your pages send to different crawlers. Request a page under several crawler names and read the source. Ask what the bot does when it must refuse a request. A refusal that includes the regulator route keeps the customer's options open.
Internet Archive copies still show these notes, so fixing a page does not erase its history. A machine-only layer is still part of what you told the public.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Redact (redact.dev).
