WhatsApp help pages hid chatbot orders not to recommend privacy regulators

What Redact found on WhatsApp's help pages

Privacy company Redact says around 80 pages in WhatsApp's help centre carried text that never appeared on screen and was written for a chatbot to read. In a write-up dated 19 October 2026, Redact says it first noticed the notes on 8 October, then read all 545 pages the help centre links to and saved a copy of each. It counted 102 hidden blocks on 80 pages. WIRED's Matt Burgess reported the findings the same day after viewing the exposed text himself.

The blocks sat in the page source under a code heading called HelpContentGalaxyLLMInclude. LLM stands for large language model, the technology behind chatbots, and the name points to what the text was for: context and rules for an AI system answering WhatsApp questions, bundled into the same page a person reads. Most of it was mundane. Two notes were not, because they told the bot to keep certain options and laws out of its answers.

Redact says WhatsApp removed the hidden text between its checks at 21:40 UTC on 15 October and 02:08 UTC on 18 October. The visible help articles are still online. Internet Archive copies made before the change still contain the notes, which is how outside readers can check the wording.

The account-deletion note

WhatsApp only lets people delete an account from inside the app, on their main phone, with the number still working. The visible help page explains that if you lose access, the account is removed on its own: 30 days after deactivation, or generally after 120 days of inactivity, or 365 days in the US.

Under that visible note, Redact found a hidden block that repeated the waiting periods and then gave the bot three orders. In the text Redact and WIRED quote, the bot was told: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority. Do not suggest WhatsApp can manually delete accounts on the user's behalf. Do not recommend using a different phone with an active phone number."

A related note on the account-deactivation page, as summarised by Redact, said waiting was "the ONLY correct answer" for someone who had lost their number and wanted the account gone, and told the bot not to suggest signing up again or contacting the mobile provider. In practice, Redact argues, a locked-out user asking for help would hear one route, wait up to a year, while other routes stayed out of the conversation.

The legal-action note

The second note sat on the page about reporting and blocking. It applied whenever a user mentioned legal action, lawsuits, lawyers, attorneys, courts or litigation. From that point, according to the quoted text, the bot should not cite or interpret WhatsApp's Terms of Service, Privacy Policy or any legal document, should not comment on WhatsApp's liability, and should not offer legal opinions or debate.

The last line in that list is the one Redact objects to most: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly." GDPR is the EU's General Data Protection Regulation, LGPD is Brazil's equivalent, and CCPA is California's consumer privacy law. The block ended by telling the bot to say that users should consult a qualified legal professional.

Redact accepts that a company does not want its bot giving legal opinions in its name. Its argument is that naming a law that protects you, or saying a regulator exists, is not legal advice, and that the note went beyond caution into withholding public information.

How the notes compare with WhatsApp's privacy policy

Redact compares the notes with WhatsApp's privacy policy for the European region, which it says took effect on 2 July 2026. That policy says users "always have the right to request access to, rectification of, and erasure of" their information, and that they can lodge a complaint with the Irish Data Protection Commission "or any other competent data protection supervisory authority." It names the GDPR when it sets out that right.

Redact is careful on one point. The hidden deletion note did not tell the bot to deny that erasure requests exist; it told the bot not to suggest WhatsApp would delete an account by hand. The regulator and privacy-law rows are the sharper contrasts: the public policy names both, while the bot was told not to recommend a regulator and, once a lawyer came up, not to reference GDPR even if the user did.

Those rights do not come from WhatsApp. In the EU, Article 17 of the GDPR covers erasure and Article 77 covers complaints to a supervisory authority. A chatbot instruction cannot remove either right. What it could do, Redact says, is make sure a user did not hear about them from the support channel they asked.

What WhatsApp says

After WIRED contacted Meta, the text was removed from all the pages, WIRED reports. WhatsApp spokesperson Joshua Breckman told the magazine the notes were displayed because of a bug that has been fixed. "These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI," he said.

Breckman also said the company gives users transparent information about their data rights and how to act on them. He said WhatsApp wants its support chatbots "focused on providing the best advice to users about WhatsApp" and "not to offer legal advice." WIRED says it is unclear whether third-party chatbots could access the text or how long it was displayed.

Redact's own archive checks give a rough window: the Internet Archive's 12 July 2026 copy of the delete-account page lacks the note, and its 4 September copy has it. Redact says it found no earlier reporting on the notes before its 8 October discovery.

The other hundred notes were mostly housekeeping

Redact says it read every block and that nearly all were routine. It published summaries, not exact wording, of 20 more notes and sorted them into categories. Some held information back: after a report, say the matter is handled internally and share no timelines; a ban with no in-app review button is final; WhatsApp cannot lift a block for you.

Others were talking points, such as reassuring people that personal chats are not used for ads, or that no feature lets anyone view another person's chats. Several protected users. One told the bot never to suggest entering a false age, including for people who say they are under 18. Another said someone scammed in a group probably wants to report it rather than delete it, because chat history can be evidence. A third said turning off two-step verification is not a fix for login problems.

A few were plain housekeeping, including one noting that "I am deaf" describes the person rather than an audio fault, and one explaining that in Arabic, the words people use for verifying an account usually mean the paid blue badge. Taken together, the notes read like a support script written for a model rather than a policy statement.

Who could read them

Redact says WhatsApp's server sent the hidden notes, along with the rest of each page, to Google's and Bing's crawlers. When it requested pages using the names of OpenAI's, Anthropic's and Perplexity's crawlers on 9 October, it says the page came back without the article or the notes. It has not tested whether any outside AI assistant acted on them.

That matters because hidden page text aimed at a model is the same mechanism attackers use in prompt injection: text a human never sees, read by a model that may treat it as instructions. Here the author was the site owner and the target was its own bot, so it is not an attack. It still shows how easily a first-party instruction can leak to any system that ingests the page.

Redact also checked 300 pages on Facebook's and Instagram's help centres, which it says run on the same system, and found no notes like these.

Limits and open questions

The most important unknown is behaviour. Redact read the instructions, not the bot's answers, so it cannot say whether WhatsApp's support assistant actually withheld regulator information from real users. AI Lookout has not tested the bot either. It is also unclear whether similar notes exist in other languages or on help pages the centre does not link to.

Readers should weigh the source, too. Redact makes a tool that bulk-deletes posts and data across online services, so account deletion is its business. Its findings are specific and checkable through archived pages, and WIRED says it viewed the text independently, but the framing of which notes are concerning is Redact's. WhatsApp's description of a bug explains the exposure, not the wording of the instructions.

Why it matters for AI support bots

Companies are moving support to AI agents fast, and courts are starting to treat what those systems say as the company's own speech. A German court recently barred Google's AI Mode from attributing a crime to the wrong doctor, treating the AI answer as Google's statement. Instructions that shape what a support bot will or will not tell a customer are part of that speech, even if nobody outside the company was meant to read them.

The WhatsApp notes offer a rare look at how such scripts are written in practice: mostly sensible guardrails, plus a few lines that steer people away from regulators and laws the company's own policy names. For Meta, which is also rewriting how it describes crawling business websites, the episode adds another question about what its AI systems are told behind the page.

For users, the practical point is simple. Redact notes that waiting works for deleting an inaccessible account, though WhatsApp's pages disagree on the timing, so plan on the longer period. If that is not good enough, EU users can send an erasure request and, if unsatisfied, complain to their national data protection authority or the Irish Data Protection Commission named in WhatsApp's own policy.