WhatsApp Help Pages Exposed Hidden Instructions for Its AI Support Chatbot

Hidden instructions in around 80 WhatsApp help and FAQ pages offered a rare look at how Meta directs its AI support chatbot. The text appeared in page code rather than in the answers shown to readers, and included guidance on account deletion, accessibility, legal questions and features that weren’t yet generally available.

The discovery, reported by privacy company Redact, involved about 100 blocks of text written for a large language model. The instructions were embedded in pages on WhatsApp’s help site, including content about account deletion and reporting or blocking other users. The pages gave readers practical support information; the hidden material supplied additional rules for how the chatbot should respond to questions about that information.

That distinction matters. A help page tells a person what the service says about a topic. Instructions for a chatbot shape how an automated system handles questions, including which subjects it should avoid and when it should direct someone elsewhere. The exposed text made that second layer visible, even though it wasn’t meant to appear to the public.

Instructions for sensitive questions

Some of the clearest examples concerned requests touching on legal matters. On a page about reporting and blocking, the chatbot was told how to respond if a user mentioned lawsuits, lawyers, courts or similar subjects. The directions said not to interpret WhatsApp’s terms or privacy documents, make claims about the company’s liability, or provide legal opinions on its behalf.

Other instructions told the chatbot not to interpret privacy rules named by users. A separate passage on the account deletion page directed it not to suggest that users contact outside data protection bodies about their concerns. Taken together, the rules appear designed to keep the chatbot focused on product support and away from answers that could be mistaken for the company’s legal position.

WhatsApp spokesperson Joshua Breckman said the chatbot is intended to help users with questions about the service, rather than offer legal advice. He also said the company provides information about users’ data rights and ways to act on them. That explanation describes the intended role of the system; the exposed text showed how some of its boundaries were expressed to the model.

The boundary can be useful for a support system: automated answers about complex disputes can confuse users if they sound authoritative without resolving the underlying issue. But a restriction on what a chatbot may discuss can also affect what users hear when they ask a question. The pages revealed the instruction itself, not how consistently the chatbot followed it or what responses users received in practice.

More than a list of forbidden answers

The hidden material also covered situations that aren’t legal or account questions. One instruction said that if a user mentioned being deaf or hard of hearing, the chatbot should treat that as background about the person rather than assume it was the support request. Another directed it not to mention community features that weren’t broadly available yet.

These examples show the range of decisions packed into chatbot guidance. A model needs to distinguish personal context from the task a person wants help with, and it needs to avoid presenting unreleased functions as available. Those directions don’t establish how the system performs on those tasks, but they do show what its creators wanted it to take into account.

Dan Saltman, Redact’s founder and CEO, said the gap between what people see and what a language model is told stood out. Redact regularly scans large technology companies’ sites and privacy materials, he said. Many of the WhatsApp directions were ordinary support guidance that pointed people to other help options, but their presence in public page code appeared unintended.

A page-code exposure, now removed

The instructions appeared under code blocks labeled “HelpContentGalaxyLLMInclude.” Redact found them in the underlying code of WhatsApp pages, and archived versions could still show the text after it was removed from the current pages. After WIRED contacted Meta, the company removed the material from the pages. Breckman said a bug had caused the chatbot instructions to be displayed and that the issue was fixed.

The account leaves several practical questions open. The available information doesn’t establish how long the text was exposed, whether other automated services could retrieve it, or whether users saw chatbot responses shaped by every instruction. It also doesn’t show whether the guidance was part of the chatbot’s only source of direction; the exposed blocks are a glimpse at one layer, not a full account of the system.

For teams publishing pages that also feed automated support tools, the incident illustrates a technical boundary worth checking: material intended to guide a model can wind up in the same public code as material intended for a reader. The exposure didn’t reveal the chatbot’s full operation, but it did make its response rules inspectable. The useful next question is whether those rules produce clear, accurate help when users ask about sensitive or ambiguous problems.