Meta's WhatsApp inadvertently exposed hidden instructions for its chatbot through help pages on its website. These pages, which are meant to answer user questions, contained text meant for a large language model, not for public viewing. The instructions, found in the code of the webpages, included guidance on steering the chatbot away from providing unhelpful or legal advice.
The exposed text, according to an analysis by privacy company Redact, appeared in the code of around 80 help and FAQ pages.
These instructions provided insight into how chatbots are trained, including guidance on what not to say to users and how to avoid legal interpretations.
One example was a section on deleting an account that instructed the chatbot not to recommend contacting any regulatory body.
The instructions were hidden under code blocks with the heading 'HelpContentGalaxyLLMInclude' and were not visible on the live pages.
When WIRED contacted Meta about the issue, the text was removed from all pages, with WhatsApp spokesperson Joshua Breckman stating it was due to a bug that has now been fixed.
He emphasized that the instructions were for the chatbot, not for users, and were meant to ensure accurate responses.
"The stand-out thing here is the gap between what people are told vs LLMs," said Dan Saltman, CEO and founder of Redact.dev. He noted that many of the instructions were straightforward, directing users to other support channels, but they were not meant for public exposure. The company regularly scans the websites of big tech firms for such content.
The exposed instructions also included guidance on how to respond if a user mentions legal action, lawsuits, or data protection regulations. The text advised the chatbot not to reference legal documents or offer interpretations on WhatsApp's liability. Another section told the chatbot not to mention specific upcoming features that are not yet available.
Meta did not say how long the text had been displayed or whether it could be accessed by third-party chatbots.
Breckman reiterated that the company provides users with transparent information about their data rights and that the chatbots are focused on offering the best advice about WhatsApp, not legal counsel. The company emphasized that the text was removed following the bug fix.
Source: wired
