WhatsApp help pages leak hidden instructions for Meta AI chatbot

A technical bug on approximately 80 WhatsApp help and FAQ pages inadvertently exposed hidden instructions intended for Meta’s large language model (LLM) support chatbot. The text, discovered by the privacy firm Redact, was embedded in the website's underlying code and was not visible to standard users.

The leaked instructions included a section titled “WHAT NOT TO SAY,” which provided specific guidance on how the AI should handle sensitive inquiries. Notably, the chatbot was instructed not to recommend that users contact data protection authorities, regulators, or supervisory bodies, even in cases involving account deletion. Additionally, the instructions directed the AI to avoid citing or interpreting privacy laws—such as Europe’s GDPR, Brazil’s LGPD, or California’s privacy laws—once a user mentioned legal action, lawyers, or litigation.

Other instructions covered how to respond to users with hearing impairments and prohibited the chatbot from mentioning features not yet released to the public. Meta has since removed the text from the affected pages, stating that the exposure was due to a bug and that the instructions were intended to ensure the chatbot provided accurate information rather than legal advice.