Privacy research firm Redact examined the WhatsApp support site and identified roughly 80 help and FAQ pages that contained hidden blocks of text intended for an internal large-language model. The analysis uncovered about 100 separate code fragments that were not displayed to users but were embedded in the page source. These fragments appeared under a heading labelled “HelpContentGalaxyLLMInclude” and were only visible when viewing the underlying HTML.
The hidden instructions included sections titled “WHAT NOT TO SAY” and explicit directives to avoid providing legal counsel. One snippet on the account-deletion page warned the model not to suggest contacting any data-protection authority, regulator, or supervisory body. Additional guidance prohibited the chatbot from referencing or interpreting privacy regulations such as GDPR, LGPD, or CCPA, even when users raised those topics directly.
According to the code, the directives were meant to steer the language model away from unhelpful or risky answers and to channel users toward official support channels. The instructions emphasized delivering accurate, relevant responses for people who message WhatsApp Support with AI assistance, while deliberately limiting the bot’s ability to give legal advice or cite regulatory frameworks. The presence of a specific heading suggests the text was part of a systematic prompt-engineering process for Meta’s own chatbot.
Redact’s CEO Dan Saltman noted a clear gap between the information presented to users and the internal guidance fed to the LLM. He said his firm routinely scans the privacy policies and site content of major technology companies, and that many of the WhatsApp directives simply redirected users to other official support resources. Saltman stressed that the hidden prompts were not intended for public consumption.
After WIRED reported the issue, Meta removed the hidden blocks and described the exposure as a bug that has been fixed. WhatsApp spokesperson Joshua Breckman said the text was intended solely for the chatbot to ensure accurate and relevant answers, not to address users directly. He added that the company trains its support bots to be as helpful as possible, but did not clarify whether third-party bots could have accessed the instructions or how long they were visible.
The hidden instructions are no longer present in the live WhatsApp help pages, but the archived source code still shows the fragments. Meta maintains that it provides transparent information about data rights and aims to keep its support chatbots focused on product advice rather than legal guidance. The incident highlights the challenges of managing internal LLM prompts and the risk of unintentionally exposing them through web content.