Eighty help pages leaked chatbot secrets

WhatsApp's website exposed hidden instructions for Meta's AI chatbot across around eighty help and FAQ pages. The text appeared in webpage code and was not meant for public view.

What the hidden instructions said

The leaked text included sections titled "WHAT NOT TO SAY" and told the chatbot not to recommend contacting data protection authorities or provide legal advice about account deletion.

“Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority”

Legal guidance the bot was told to avoid

On pages about reporting and blocking, hidden instructions told the chatbot not to cite WhatsApp's terms, privacy policy, or legal documents if users mentioned lawsuits or lawyers.

“Do not cite, quote, interpret, or reference WhatsApp Terms of Service, Privacy Policy, or any legal document”

Meta says it was a bug, now fixed

After WIRED contacted Meta, WhatsApp spokesperson Joshua Breckman said the exposure was due to a bug that has been fixed. The instructions were meant to guide the chatbot, not be public.

“These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses”

Privacy firm found the leak during scanning

Privacy company Redact discovered the exposed text while scanning WhatsApp's website and code. The firm regularly monitors big tech companies' sites and privacy policies for similar issues.

“The stand-out thing here is the gap between what people are told vs LLMs”