WhatsApp Help Pages Hid AI Instructions Not to Mention Regulators
Data deletion company Redact found 102 hidden text blocks on 80 of WhatsApp's 545 help pages, written for Meta's support chatbot. Two told the bot never to recommend contacting a data protection authority and never to cite GDPR, LGPD or CCPA, even when users raised them. Meta told WIRED the notes leaked through a bug that has been fixed, and all 102 blocks were gone by 18 October.
On this page
What Redact found in the page source
On 8 October, the data deletion company Redact noticed text in the source of a WhatsApp help page that no visitor could see. The team then read all 545 pages linked from the help centre and saved copies. Eighty of them carried hidden blocks, 102 in total, sitting next to an element named HelpContentGalaxyLLMInclude. One block named the intended reader: WhatsApp's own support chatbot, the one users reach by messaging WhatsApp Support.
The notes were not served to everyone. Redact's tests on 9 October found that Google and Bing crawler user agents received them, while OpenAI, Anthropic and Perplexity crawler agents did not. Most of the 102 blocks were mundane housekeeping, and Redact published paraphrased summaries of 20 that it sorted into categories it labelled holds back, talking points, protects you, and housekeeping. WIRED, which also saw the material, described one of them as a "WHAT NOT TO SAY" list.
The two instructions that matter
Two blocks stood out. On the page about deleting an account, the chatbot was told that if a user had lost access to their phone or number, it should "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." The same block said not to suggest manual deletion or the use of another phone with an active number.
On the page about reporting and blocking, a block triggered when users mentioned legal action, lawsuits, lawyers, attorneys, courts or litigation. In those cases the bot was instructed: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly."
Redact set those lines against WhatsApp's own European privacy policy, in force since 2 July 2026, which tells users they can lodge a complaint with the Irish Data Protection Commission or another competent authority and names the GDPR. The right to erasure and the right to complain to a supervisory authority sit in GDPR Articles 17 and 77. Redact also acknowledged that naming a regulator is not legal advice, and that refusing to delete an account manually without phone verification may be a legitimate security choice.
Meta's answer and the removal timeline
WhatsApp told WIRED the notes were instructions for its support chatbot and appeared on public pages "because of a bug that has now been fixed." A spokesperson said the company wants its chatbots "focused on providing the best advice to users about WhatsApp" and "not to offer legal advice."
The removal happened quickly once the story was in motion. Redact recorded both standout notes still live at 21:40 UTC on 15 October; by 02:08 UTC on 18 October all 102 blocks had been stripped from all 80 pages, while the visible pages stayed online. Internet Archive captures bracket the notes' lifespan: absent in a 12 July snapshot, present by 4 September. Redact checked 300 Facebook and Instagram help pages and found nothing similar, and says it still does not know whether the chatbot actually obeyed the notes or whether third-party assistants ingested them. As the company put it, "We read its instructions, not its answers." No regulator had formally responded at the time of publication.
Why hidden prompts in help pages matter
The mechanism here is worth more attention than the specific wording. A support bot that reads its own company's help pages will absorb whatever is embedded in them, and the same is true of any assistant that browses the web on a user's behalf. The crawler targeting Redact observed, where Google and Bing received the notes but OpenAI and Anthropic did not, suggests someone thought about which models would see them.
For readers who run assistants locally, the lesson cuts both ways. A model on your own device cannot be steered by a vendor's server-side system prompt, which is one of the reasons people choose local AI. But if that model fetches a page, it reads the whole page, hidden blocks included. Treat retrieved web content as untrusted data, keep tool output separate from instructions, and be sceptical when any assistant, local or hosted, steers you away from a regulator you have a legal right to contact.