WhatsApp Help Pages Exposed Hidden Instructions for Its AI Support Chatbot

Hidden instructions in roughly 80 WhatsApp help pages offered a rare look at how Meta directs its chatbot to answer support questions. The text was written for a large language model, not for people browsing the pages, and included rules about legal topics, accessibility statements and features that weren’t broadly available.

The instructions appeared in the pages’ underlying code, according to an analysis by privacy company Redact. The company identified about 100 blocks of text. They were associated with WhatsApp’s help and FAQ pages, where they supplied the chatbot with context beyond what visitors could see in the page itself.

That distinction matters for anyone assessing an AI support tool. A help page presents information to a reader; instructions embedded in its code can also shape how a chatbot uses that information. In this case, the text spelled out subjects the model should avoid and how it should handle certain user messages. It offers a small but concrete example of how product teams try to limit chatbot responses through written directions.

Instructions beneath the help pages

The hidden blocks appeared under the heading “HelpContentGalaxyLLMInclude.” One example was attached to a page about deleting a WhatsApp account. Among other directions, it told the chatbot: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” about the issue.

On a separate page about reporting and blocking, the instructions addressed questions involving legal action, lawyers, courts and litigation. The chatbot was told not to cite or interpret WhatsApp’s terms, privacy policy or other legal documents, and not to make claims about the company’s liability. Other text instructed it not to interpret privacy rules, including GDPR, LGPD and CCPA, even when users raised those topics.

These directions set boundaries for the chatbot’s replies; they don’t establish how it would respond in every conversation. The pages revealed the written guidance, but the available findings don’t say how often people encountered the chatbot, what answers it produced, or whether other chatbots could access the exposed text.

Some instructions dealt with recognizing the user’s immediate need. If someone said they were deaf or hard of hearing, the model was told to treat that statement as background rather than as the support request itself. Another direction said not to mention community features that weren’t yet generally available.

Meta says a bug exposed the text

Dan Saltman, Redact’s CEO and founder, said his company regularly scans large technology companies’ websites and privacy pages. He described the gap between what visitors see and what language models are instructed to do as the most striking part of the discovery. Saltman also said many of the directions were straightforward and pointed users toward other support options, while appearing unintended for public view.

After WIRED contacted Meta, the text was removed from the live pages. WhatsApp spokesperson Joshua Breckman said a bug had caused the instructions to appear and that the issue was fixed. He said the directions were meant for WhatsApp’s support chatbot, for people who choose to message WhatsApp Support using AI, and were intended to help it provide relevant and accurate answers. Archived versions of the pages still showed the text in their code.

Breckman said the company aims to make its support chatbots helpful, and that WhatsApp wants them to focus on product support rather than legal advice. He also said the company provides information about users’ data rights and actions they can take. The exposed instructions themselves, however, specifically told the chatbot not to interpret privacy rules or WhatsApp’s legal documents.

What the exposure reveals—and what it doesn’t

The episode shows how instructions for an AI support system can sit alongside content written for human readers. That arrangement can help a chatbot tailor replies to a particular help page, while also creating a second layer of material that may not be intended for visitors. Here, a website bug made that layer visible in page code.

There are clear limits to what can be concluded from the exposed text. It doesn’t show whether the chatbot followed every instruction, how those directions were tested, or how long they were present. Meta said the bug had been fixed, but the available account doesn’t establish whether third-party chatbot systems could read the text while it was live.

For teams building AI support, the practical takeaway is that chatbot instructions and public help content need to be handled as distinct parts of a product. Both can affect what users receive, and a mistake in how a page is assembled can expose material intended for the model. WhatsApp’s hidden text makes that boundary visible; whether the chatbot follows the directions in real support conversations remains unanswered.

Sources: WIRED’s Meta coverage, WhatsApp’s help pages, and Redact’s analysis.