Around 80 WhatsApp help and FAQ pages exposed hidden instructions intended for a Meta chatbot, according to an analysis by privacy company Redact reported by WIRED. Roughly 100 blocks of text appeared in page code and included sections labeled “WHAT NOT TO SAY,” along with directions not to provide legal advice.
One instruction attached to an account-deletion page told the model not to recommend contacting a data-protection authority or regulator. Much of the other guidance directed users toward official support channels. The material offered an unusual public view of how a company constrains a support chatbot, but it was not presented to ordinary readers as part of the visible help article.
Meta removed the text after WIRED contacted the company. A WhatsApp spokesperson said a bug displayed information that was not intended to be public. The exposure does not show that the chatbot followed every instruction or that user messages were leaked. It does show that model prompts and policy guidance can become part of a product’s attack and disclosure surface. Teams deploying support agents need to assume hidden instructions may be inspected and avoid placing secrets in them.