Hidden WhatsApp help page instructions told AI assistants to withhold regulator and GDPR details

Invisible instructions in the help centre

On October 8, 2026, researchers at redact.dev discovered text embedded in WhatsApp's help pages that never rendered on screen but was served to software reading the page — and parts of it told AI assistants to keep useful information away from users.

After the discovery, the team read all 545 pages linked from the help centre and saved a copy of each. Eighty pages carried hidden AI-facing text, roughly 102 blocks in total. Most were benign operational notes, but two stood out.

The first sat beneath the visible note on the account deletion page. WhatsApp only allows deletion from inside the app, on a working phone with a working number, so locked-out users must wait for automatic removal — 30 days after deactivation, or 120 days of inactivity (365 in the US). The hidden text repeated those periods and added orders for the AI: do not suggest WhatsApp can delete an account manually, and do not recommend contacting any data protection authority, regulator or supervisory authority. A user who had lost their phone and wanted out was steered toward waiting up to a year, with no other route offered.

The second block, on the page about reporting and blocking, activated the moment a user mentioned a lawyer, a lawsuit, legal action or a court. From then on the AI was told to avoid anything legal, including an instruction not to reference, cite or interpret the GDPR, Brazil's LGPD or California's CCPA — including cases where the user was the one citing the statute.

At odds with WhatsApp's own privacy policy

The hidden instructions contradicted documents WhatsApp is obliged to publish. Its European privacy policy, effective July 2, 2026, says users can always request access, rectification and erasure of their information, and may complain to the Irish Data Protection Commission or any other competent supervisory authority. The policy also names the GDPR when setting out the right to complain.

As redact.dev points out, those rights stem from Articles 17 and 77 of the GDPR, and an instruction to a chatbot cannot remove them. What it could do was ensure users never heard about them from the channel meant to help them.

The researchers acknowledge part of the picture is defensible. A company should not want its chatbot issuing legal opinions in its name, and directing people to a lawyer is sound practice. But naming a law that protects a user, or noting that a regulator exists, is not legal advice — the privacy policy already does both. Similarly, WhatsApp may genuinely be unable to verify who owns an account without the phone number, which is a fair reason to refuse manual deletion, but it is not a reason to leave the regulator out of the answer.

Who received the text

The hidden blocks carried an identifier containing the letters LLM — HelpContentGalaxyLLMInclude — and WhatsApp told WIRED the notes were instructions for its own support chatbot. The server also sent them, along with the rest of each page, to Google's and Bing's crawlers. When redact.dev requested the pages on October 9 using the crawler names that OpenAI, Anthropic and Perplexity use, they came back without the notes. Facebook's and Instagram's help centres run on the same platform; a check of 300 of those pages found no comparable instructions.

Archived snapshots narrow the timeline. An Internet Archive copy of the deletion page from July 12, 2026 shows no note, a September 4 copy includes it, and WhatsApp had removed the notes by an October 18 capture. redact.dev found no earlier reporting on them.

What is still unknown

The researchers read the chatbot's instructions, not its answers, so it remains unknown whether the bot actually followed them, or whether any outside AI assistant acted on the text when it was served. The full lifespan of the notes, possible versions in other languages, and pages the help centre does not link to are also open questions. The blog publishes a table of twenty further hidden blocks, most of them milder relations of the two above. One small inconsistency also surfaced: WhatsApp's own pages disagree on how long an unused account survives, so the researchers suggest planning for the longer window.

Why it matters

Content that reads differently to machines than to people is a new lever over what users learn. As more people reach answers through AI intermediaries, help pages can be tuned invisibly — steering conversations away from regulators, statutes and remedies while the visible page, and the published privacy policy, tell a different story. That divergence is the core issue: a company's public commitments and its machine-facing behaviour can split apart without any human noticing, and only crawler-level auditing catches it. Checking what a page says, and to whom, is becoming a basic step in judging whether a service is being straight with its users.