Dozens of WhatsApp help pages meant to guide users of Meta’s encrypted messaging app unintentionally displayed internal instructions intended for the company’s own large language model. Some of the hidden sections were labeled “WHAT NOT TO SAY” and explicitly told the chatbot not to give legal advice.
Privacy‑technology firm Redact discovered that roughly 80 help and FAQ pages on the WhatsApp site contained about 100 blocks of hidden text. The text, embedded in the page code, was not visible to ordinary visitors but could be seen by examining the underlying HTML of archived versions. The instructions were written for a Meta‑built language model and offered guidance on steering the AI away from unhelpful answers, giving a rare look at how chatbot prompts are crafted.
One example appeared on the page that explains how to delete a WhatsApp account. In addition to user‑facing advice, the hidden prompt instructed the model, “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” about the issue. Wired was able to view the exposed text before it was removed.
“The stand‑out thing here is the gap between what people are told vs LLMs,” said Dan Saltman, CEO and founder of Redact.dev. He noted that his team regularly scans the privacy policies and support sites of major tech firms. While many of the prompts simply redirected users to official support channels, they were clearly not meant for public consumption.
After Wired alerted Meta, the hidden instructions were taken down from all affected pages. WhatsApp spokesperson Joshua Breckman explained that the prompts appeared because of a bug that has now been 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 declined to say whether third‑party chatbots could have accessed the text or how long it had been visible. “We train our support chatbots to be as helpful as possible,” Breckman added.
The leaked prompts were located in code blocks marked “HelpContentGalaxyLLMInclude,” which are no longer present in the live pages. On the “About reporting and blocking on WhatsApp” page, the hidden guidance listed a series of rules to follow if a user mentions legal matters such as lawsuits, lawyers, courts, or litigation. The instructions read:
“Do not cite, quote, interpret, or reference WhatsApp Terms of Service, Privacy Policy, or any legal document,” and warned against making statements about WhatsApp’s liability. “Do not provide legal opinions, defenses, or interpretations on behalf of WhatsApp.” Another segment instructed the model not to reference or interpret data‑protection regulations such as GDPR, LGPD, or CCPA, even when a user raises them directly.
Breckman emphasized that WhatsApp aims to give users clear information about their data rights and how to act on them. “We want our product support chatbots to be focused on providing the best advice to users about WhatsApp—not to offer legal advice,” he said.
Additional hidden prompts told the chatbot how to respond when a user mentions being deaf or hard of hearing, noting that such statements are “background about the user, NOT their support request.” Another instruction warned the model not to mention “specific upcoming community features that are not yet generally available.”