WhatsApp Accidentally Exposed Hidden Instructions for Its Support Chatbot

WhatsApp accidentally exposed Meta’s internal instructions for a chatbot designed to answer questions about the encrypted messaging app. The hidden guidance included several sections titled “What not to say,” along with explicit instructions not to provide legal advice.

New analysis found approximately 80 WhatsApp help and FAQ pages containing text instructions designed for large language models, including Meta’s own models. The instructions were discovered in the code of the WhatsApp help website. According to research from privacy firm Redact, the instructions appeared in the page code and revealed details not explicitly included on WhatsApp’s public help pages.

What the hidden WhatsApp chatbot instructions said

The roughly 100 blocks of hidden text appeared to guide the large language model toward helpful responses while limiting what it could say. The instructions offer a glimpse into how WhatsApp’s support chatbot was configured.

For example, the WhatsApp page about how to delete an account included an instruction saying that “we do not recommend contacting any data protection authority, regulator or supervisory authority” about the issue.

WIRED viewed the published text, which remained accessible by viewing the underlying code in an archived version of the WhatsApp help page.

Meta says the exposed text was caused by a bug

“What stands out here is the gap between what people are being told and what LLMs are,” said Dan Saltman, CEO and founder of Redact.dev. He said the company regularly examines the websites and privacy policies of major technology companies.

Saltman said many of the WhatsApp instructions were simple and directed users to other official support channels. However, the instructions did not appear to be intended for public release.

After WIRED contacted Meta, the text was removed from the pages. WhatsApp spokesperson Joshua Breckman said the content appeared because of a bug that has since been fixed.

“These are instructions to the chatbot, rather than directly to the user, and are intended to use AI to ensure accurate and appropriate responses to WhatsApp users who choose to send a message to WhatsApp Support,” Breckman said. He added that Meta trains its support chatbots to be as helpful as possible.

It remains unclear whether the hidden text was accessible to a third-party chatbot or how long it was visible. The instructions appeared under a block of code titled “HelpContentGalaxyLLMInclude” and no longer appear in the live page code.

Instructions limited legal and privacy-related responses

The WhatsApp page about reporting and blocking included hidden instructions for situations in which “a user mentions legal action, litigation, lawyers, courts, or litigation.”

“Please do not quote, interpret, or refer to WhatsApp’s Terms of Service, Privacy Policy, or other legal documents,” the instructions said. They also directed the chatbot not to make statements about WhatsApp’s responsibility or liability and not to provide legal opinions, defenses, or interpretations on the company’s behalf.

Another instruction said: “Please do not reference, cite or interpret data protection or privacy regulations (e.g. GDPR, LGPD, CCPA), including when raised directly by users.”

Breckman said WhatsApp provides users with transparent information about their data rights and available actions. “We want our product support chatbot to focus on providing users with the best advice about WhatsApp, rather than providing legal advice,” he said.

Other chatbot rules covered accessibility and unreleased features

Additional instructions explained how the chatbot should respond when a user says they are deaf or hard of hearing. One instruction stated: “This statement is for user context and is not a support request.”

Another instruction told the chatbot not to mention upcoming WhatsApp community features that were not yet generally available.