A bug on WhatsApp’s website exposed hidden instructions that Meta gives its AI chatbot. The instructions appeared on around 80 help and FAQ pages and included rules on what the chatbot should and should not tell users. Some sections were titled “WHAT NOT TO SAY”. They also told the chatbot not to provide legal advice or discuss privacy laws. According to a report by Wired, citing analysis by privacy company Redact, the text was meant to guide Meta’s chatbot when answering support questions. However, the instructions appeared in the website’s code, where they could be viewed. Meta removed the text after Wired reported the issue.
The exposed text included about 100 blocks of instructions for the chatbot. These were found in the code of WhatsApp’s help pages, instead of in the information shown directly to users. The instructions were meant to help the chatbot respond to questions.
On the account deletion page, the chatbot was told: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” about the issue.
Other instructions explained how the chatbot should respond when users mention lawsuits, lawyers or legal action. It was told not to interpret WhatsApp’s legal documents or discuss whether the company could be held responsible.
The instructions also said: “Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly.”
The company has now removed the exposed text from its website. WhatsApp spokesperson Joshua Breckman told Wired that the problem was caused by a bug that has now been fixed.
“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 was quoted as saying in the report.
He added that the company wants its support chatbots to focus on helping users with WhatsApp, rather than providing legal advice.
The instructions appeared under a code heading called “HelpContentGalaxyLLMInclude”. They are no longer visible in the live pages’ code. However, Wired found that they could still be viewed in archived versions of the pages.
It remains unclear how long the instructions were exposed or whether third-party chatbots could access them.