WhatsApp accidentally exposed hidden instructions meant for Meta’s AI support chatbot across dozens of help pages, revealing specific directions on what it should not tell users about legal action, privacy regulations and upcoming WhatsApp features.
According to analysis by privacy company Redact, around 80 pages on WhatsApp’s website contained instructions for a large language model (LLM). These directions were not displayed as part of the normal help articles but could be found in the underlying code of the webpages.
Redact reportedly discovered nearly 100 blocks of hidden text across WhatsApp’s help pages. Some instructions were labelled “WHAT NOT TO SAY” and appeared to tell the AI chatbot how to handle particular questions or situations. The instructions appeared in code blocks carrying the header “HelpContentGalaxyLLMInclude.”
Dan Saltman, CEO and founder of Redact.dev, told Wired that many of these instructions were relatively straightforward and directed users towards official support options. However, the material did not seem intended for public viewing.
“The stand-out thing here is the gap between what people are told vs LLMs,” Saltman said. Some of the most detailed instructions concerned users who raised legal issues while seeking WhatsApp support.
After Meta was contacted about the exposed material, the instructions were removed from the affected WhatsApp pages. The “HelpContentGalaxyLLMInclude” text is no longer visible in the code of the live pages.
