Problem Solutions
WhatsApp's Meta AI Hidden Instructions Leaked: What It Is Told Not to Say
The instructions were never meant to be public. For an unknown period, internal guidance written for the AI chatbot behind WhatsApp Support sat inside the code of WhatsApp's own help pages, readable by anyone who viewed the page source. The privacy company Redact found the text while scanning the help center; after WIRED contacted Meta, it was removed. Among the exposed notes were sections headed "WHAT NOT TO SAY," listing subjects the chatbot was told to avoid — including pointing users toward data protection regulators.
What exactly leaked
The hidden text appeared in code blocks carrying the heading "HelpContentGalaxyLLMInclude." Dan Saltman, CEO and founder of Redact, read and archived all 545 pages linked from WhatsApp's help center and counted 102 hidden blocks across roughly 80 pages. The notes were instructions for one of Meta's large language models: guidance for the AI that answers WhatsApp Support messages. Most were routine, but the text was never displayed in the help articles themselves; archived copies confirm it was live by early September and still present on October 9. After Meta was contacted, the blocks were removed; none remained by October 18. It is still unclear how long the text had been exposed, or whether third-party chatbots could have accessed it.
What the instructions told the chatbot not to say
The most detailed rules covered legal and privacy questions. On the help page explaining how to delete an account, the chatbot was told: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." The same block added that it should not suggest WhatsApp can manually delete accounts on a user's behalf. On the page about reporting and blocking users, another set applied whenever a user mentioned legal action, lawsuits, lawyers, attorneys, courts, or litigation: "Do not cite, quote, interpret, or reference WhatsApp Terms of Service, Privacy Policy, or any legal document." The chatbot was also told not to comment on whether WhatsApp could be held liable, or to offer legal opinions on the company's behalf. Another instruction barred it from referencing privacy laws at all: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly." Other notes were less sensitive: one clarified that if a user said they were deaf or hard of hearing, that was background about the user, not their support request; another told the chatbot not to mention upcoming community features that are not yet generally available.
Meta's response
WhatsApp spokesperson Joshua Breckman told WIRED the text appeared because of a bug, now 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. He added that the company wants its support chatbots to focus on advice about WhatsApp, not legal advice. Saltman, whose firm routinely scans big tech sites and privacy policies, took a different view: "The stand-out thing here is the gap between what people are told vs LLMs." Redact did not test whether the chatbot actually followed these instructions, and it examined only the English-language pages.
What this means for ordinary WhatsApp users
Nothing in the leak changes how your private chats work: the instructions governed a support chatbot, not your conversations with friends and family, which remain end-to-end encrypted, and Meta AI can only see messages deliberately sent to it. The practical lesson is still worth knowing: when you ask an AI support assistant about your rights, its answer is shaped by rules you cannot see, and it is not a neutral source on complaints, regulators, or legal options. If you believe WhatsApp has mishandled your data, you can contact your national data protection authority directly, whether or not a chatbot suggests it.
What you can do about Meta AI in your chats
Meta offers no setting that removes Meta AI from WhatsApp entirely, but you can limit its visibility and keep it out of specific conversations. First, some users now have a "Show Meta AI Button" toggle: go to Settings, then Chats, and switch it off to hide the blue Meta AI circle on your chat list. Meta is rolling this out in phases, so it may not appear on your phone yet. Second, if a Meta AI chat sits in your list, open it, tap its name at the top, mute notifications, and archive it. Third, for sensitive conversations, open the chat, tap the contact or group name at the top, and turn on Advanced Chat Privacy. With it on, Meta AI cannot be invoked in that chat, and its messages cannot be used in AI features such as @Meta AI mentions or chat summaries; it applies per chat, so enable it in each conversation you want protected. Finally, typing /reset-ai or /reset-all-ais inside a Meta AI chat deletes what the assistant remembers from your conversations.
The takeaway
The leak is less a security failure than a rare look at the rulebook behind an AI assistant. Most of the rules were unremarkable. A few — steering users away from regulators, declining to discuss privacy law even when asked — show why AI support answers are company answers, not independent advice. The text is gone from WhatsApp's pages now; the questions it raised about whose interests a support chatbot serves are not.