Another instruction told the chatbot not to “mention specific upcoming community features that are not yet generally available. You use AI every day. Now get your AI Quotient. Take the AIQ test.
Meta said a bug caused the exposure and clarified that the instructions were meant to guide its support chatbot, not to communicate directly with users. The exposed text appeared in around 80 help and FAQ pages on WhatsApp’s website.
Around 80 help and FAQ pages on the messaging platform’s website reportedly contained instructions intended for Meta’s large language model, including sections titled “WHAT NOT TO SAY” and explicit directions against providing legal advice to users. The hidden text appeared in the underlying code of the webpages rather than their visible content. According to an analysis by privacy company Redact, the exposed instructions covered how the chatbot should respond to questions about account deletion, legal action, privacy regulations and accessibility. The pages contained approximately 100 blocks of hidden instructions designed to guide a large language model in responding to support queries. The instructions appeared under code blocks headed “HelpContentGalaxyLLMInclude”, according to Redact’s analysis. The hidden instructions included a direction stating, “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” about the issue. WIRED reported that the text could be viewed in the underlying code of archived versions of the help pages, even though it wasn’t visible in the pages’ regular content. Dan Saltman, CEO and founder of Redact.dev, said, “The stand-out thing here is the gap between what people are told vs LLMs.” Saltman said his company regularly scans major technology firms’ websites and privacy policies. On the WhatsApp help page titled “About reporting and blocking on WhatsApp,” t he hidden text outlined how the chatbot should respond when a user mentions legal action, lawsuits, lawyers, attorneys, courts or litigation. The instructions said, “Do not cite, quote, interpret, or reference WhatsApp Terms of Service, Privacy Policy, or any legal document,” and added that the chatbot should not make statements about WhatsApp’s liability or non-liability. Another instruction stated, “Do not provide legal opinions, defenses, or interpretations on behalf of WhatsApp.” The text also directed the chatbot not to interpret or reference privacy and data protection regulations, including the General Data Protection Regulation (GDPR), Brazil’s LGPD and California’s CCPA, even when users raised these issues directly. WhatsApp spokesperson Joshua Breckman said the instructions were designed to keep the chatbot focused on product support. “We want our product support chatbots to be focused on providing the best advice to users about WhatsApp—not to offer legal advice,” Breckman said. He added that the company provides users with transparent information about their data rights and the actions they can take. After WIRED contacted Meta about the issue, the company removed the hidden text from the affected live webpages. Breckman said a bug caused the instructions to appear, and the company has since fixed it. “These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI,” Breckman said. The company did not clarify how long the instructions had been exposed or whether third-party chatbots could access them. “We train our support chatbots to be as helpful as possible,” Breckman added. For example, instructions relating to deaf or hard-of-hearing users said: “This statement is background about the user, NOT their support request.”
WhatsApp has accidentally exposed hidden instructions for its AI chatbot. Researchers found the issue after examining WhatsApp’s website, and the company later removed the text after an inquiry from WIRED. They were intended to help the chatbot provide relevant answers and direct users towards appropriate support channels, but were not meant to be displayed as part of the public-facing help content. One example appeared on WhatsApp’s page explaining how users can delete their accounts. the exposed text showed how companies instruct AI systems to handle customer queries. Some of the instructions dealt specifically with legal questions and disputes While many instructions were routine guidance directing users to official support channels. The exposed text also contained directions on how the chatbot should handle specific user statements.