Dozens of WhatsApp’s public help pages briefly revealed something Meta never intended users to see: the internal instructions given to the chatbot that answers support questions. Around 80 pages on the messaging service’s website exposed this behind-the-scenes guidance, and Meta has blamed a bug.
The exposed material appears to be the kind of instructions AI developers write to shape a chatbot’s behaviour before it ever reaches a user, often called a system prompt. Some of it was organised under headings labelled “WHAT NOT TO SAY,” setting out topics and responses the assistant should avoid. The instructions also told the bot not to offer people legal advice. That is a predictable rule for a company that would not want an automated support agent making statements that could later be read as legal guidance.
None of this is shocking on its own. Every major AI assistant runs on hidden rules that set its tone, limit its scope and keep it away from liability. What is unusual is seeing those rules published on the same pages the chatbot is meant to help users with. It offers a rare look at how a large tech company constrains an AI that speaks on its behalf to millions of people.
It also fits a pattern. In 2023, users pulled the hidden instructions out of Microsoft’s Bing chatbot through prompt injection, revealing its internal codename and a long list of behavioural limits. Since then, some AI firms have chosen to publish their system prompts openly, arguing that transparency beats leaks. Meta has not taken that approach, which makes an accidental exposure more awkward for it than for a company that discloses this material voluntarily.
Important details are still unclear. Meta has not said how long the instructions were visible, whether search engines indexed them, or exactly which chatbot they belonged to. It is also unknown whether the full text included anything more sensitive than tone and topic restrictions, such as how the assistant handles account security questions, which would matter more to WhatsApp’s users.
Meta has also not explained what kind of bug placed the internal text in front of the public. For a company that keeps leaning on WhatsApp’s end-to-end encryption as its main privacy selling point, mistakes in the AI layer around the app are worth watching. As Meta adds more automated assistants to its apps, the question is less whether these systems have hidden rules and more whether Meta can keep those rules, and the systems themselves, from leaking.
