WhatsApp Help Pages Leak Hidden Instructions for Meta Chatbot
A recent analysis revealed that approximately 80 web pages on WhatsApp’s website exposed internal instructions designed for a large language model chatbot. Meta stated that the text appeared due to a bug that has since been fixed.
Radwan AbdulhadiEditor and technology writer
Hidden LLM Instructions Exposed on Web Pages
Dozens of web pages on WhatsApp’s website designed to answer user questions about the encrypted messaging app inadvertently exposed the company’s internal instructions for its chatbot. According to findings from privacy company Redact, approximately 80 help and FAQ pages exposed text instructions created specifically for a large language model operated by Meta.
The hidden text appeared inside the underlying code of the web pages under code blocks featuring the heading "HelpContentGalaxyLLMInclude." These written details were not explicitly visible on the main help pages, though they contained around 100 blocks of hidden instructions aimed at steering the LLM away from unhelpful responses.
Guidance on Legal Advice and What Not to Say
Among the uncovered text were specific sections labeled "WHAT NOT TO SAY" alongside explicit directions advising the chatbot against providing legal advice to users. For example, a WhatsApp help page covering account deletion instructed the AI not to recommend contacting any data protection authority, regulator, or supervisory authority regarding the issue.
Additional restrictions appeared on the page regarding reporting and blocking on the platform. The instructions told the model how to respond if a user mentioned legal action, lawsuits, lawyers, attorneys, courts, or litigation. Specifically, the model was told not to cite, quote, interpret, or reference WhatsApp terms of service or privacy policies, nor to offer legal opinions on behalf of the company.
Industry Analysis and Redact's Discovery
Dan Saltman, CEO and founder of Redact, noted the distinct gap between the information presented to regular users and the directives given to the underlying AI models. He explained that his team regularly scans the sites and privacy policies of major technology corporations and observed that many of the uncovered instructions pointed people toward official support channels.
While archived versions of the help pages still allowed viewers to inspect the underlying code and view the exposed text, it remains unclear how long the instructions were publicly accessible or whether third-party chatbots could access them.
Meta's Response and Bug Fix
After WIRED contacted Meta regarding the exposed text, the company removed the instructions from all affected pages. WhatsApp spokesperson Joshua Breckman stated that the text was displayed due to a bug that has now been fixed.
According to Breckman, these messages served as instructions for the support chatbot rather than direct guidance for users, intended to ensure accurate responses when individuals chose to message WhatsApp Support with AI. Breckman added that the company wants product support chatbots to focus on offering proper advice about the app rather than issuing legal counsel.
Additional Chatbot Constraints
The hidden guidelines also instructed the chatbot on handling specific user statements, such as identifying when a user stated they were deaf or hard of hearing by treating the statement as background context rather than a support request. Furthermore, the instructions prohibited the chatbot from mentioning specific upcoming community features that were not yet generally available to the public.
Sources
- WIREDWhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot
Continue chronologically
Related entity coverage
Related stories
Meta Infrastructure Sets Records During 2026 World Cup
Meta details how its infrastructure handled record-breaking messaging, live video watch times, and traffic spikes during the 2026 World Cup.
Sapienz Receives ISSTA 2026 Impact Paper Award
Ke Mao, Mark Harman, and Yue Jia win the ISSTA 2026 Most Influential Paper Award for their foundational 2016 automated testing research.
Meta Brings Private Processing Infrastructure to AI Glasses
Meta expands its Private Processing confidential computing infrastructure to AI glasses, enabling secure cloud-based AI processing without exposing user data.
Apple Tightens macOS Full Disk Access Over AI Risks
Apple is updating macOS Full Disk Access controls, warning that increasingly capable AI agents create new security risks for user data.
