WhatsApp FAQ Pages Accidentally Exposed Internal AI Chatbot Instructions
Dozens of **WhatsApp** help pages inadvertently revealed internal instructions intended for **Meta's** large language model (LLM) chatbot. These hidden directives, discovered by privacy company **Redact**, included explicit 'WHAT NOT TO SAY' guidelines, such as avoiding legal advice or references to data protection authorities, offering a rare glimpse into chatbot operational policies.
A recent analysis by privacy firm **Redact** uncovered a significant oversight: approximately 80 help and FAQ pages on **WhatsApp's** official website were found to expose internal instructions meant for one of **Meta's** large language models. These directives, embedded within the webpages' code under the heading "HelpContentGalaxyLLMInclude," were not intended for public consumption.
The exposed text blocks provided specific guidance to the LLM, steering it away from unhelpful or legally sensitive responses. For instance, on the page detailing how to delete an account, the instructions explicitly stated, "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority" regarding the issue.
**Dan Saltman**, CEO and founder of **Redact.dev**, highlighted the disparity between information provided to users and the internal instructions for LLMs. While many instructions aimed to guide users to official support channels, their public exposure was clearly unintentional.
Upon being contacted by WIRED, **Meta** swiftly removed the exposed text from all affected pages. **WhatsApp** spokesperson **Joshua Breckman** attributed the display to a bug, which has since been fixed. Breckman clarified that these instructions were for their chatbot, not directly for users, and were designed to ensure accurate and relevant responses when users message **WhatsApp Support** via AI.
Further examples of the hidden instructions included protocols for handling user queries mentioning legal action. The chatbot was directed, "Do not cite, quote, interpret, or reference **WhatsApp** Terms of Service, Privacy Policy, or any legal document," and to avoid providing legal opinions or interpreting data protection regulations like **GDPR**, **LGPD**, or **CCPA**.
Breckman reiterated that **WhatsApp** aims for its product support chatbots to focus on providing product-specific advice, not legal counsel. Other directives advised the chatbot on how to respond to users identifying as deaf or hard of hearing, and cautioned against mentioning unreleased community features.