WhatsApp Exposes Internal Meta Chatbot Prompts
A code exposure across 80 WhatsApp help pages revealed Meta's internal system prompts, offering a rare look at how tech giants configure guardrails for AI customer support.
Privacy security firm Redact discovered that around 80 help and FAQ pages across WhatsApp's official website accidentally exposed system instructions intended for an internal Meta large language model. Embedded in the HTML code under the label "HelpContentGalaxyLLMInclude", the leak contained roughly 100 blocks of hidden prompt text designed to direct how the automated support system handles user inquiries. Meta spokesperson Joshua Breckman stated that the system instructions were exposed due to a bug that has since been patched, clarifying that the prompt text was meant for the chatbot rather than end users.
The leaked instructions reveal strict guardrails governing sensitive topics like legal liability and regulatory oversight. For example, on pages explaining account deletion, the instructions explicitly told the LLM, "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." On pages concerning blocking and reporting, the prompts instructed the model to avoid citing or interpreting regulations such as GDPR, LGPD, or CCPA, and forbade providing legal opinions or discussing WhatsApp's liability whenever users mentioned lawsuits or legal action.
Other directives instructed the model to avoid mentioning unreleased community features and provided guidance on contextual cues, such as noting when a user's statement about being deaf was background context rather than the main request. Redact.dev CEO Dan Saltman noted that while many directions simply aimed to route users to standard support channels, the incident reveals the distinct gap between public user documentation and the instructions given to LLMs.
For AI practitioners, the leak offers a concrete look at real-world prompt engineering and safety moderation architecture used in large-scale customer support deployments. It highlights how major platforms construct restrictive system prompts to prevent AI hallucinations, limit legal exposure, and separate user background information from actionable support queries.
This is our own summary of reporting by WIRED AI
