Meta’s AI chatbot on WhatsApp has a rulebook. For a while, anyone visiting the right corner of WhatsApp’s website could read parts of it.
Around 80 pages on the site exposed instructions that were meant for the chatbot’s eyes only, Wired reported. Meta says a bug was to blame.
The most eye-catching detail is a heading that reads “WHAT NOT TO SAY.”
What actually leaked
The exposed material was guidance written for Meta’s AI chatbot. It was not intended for the people chatting with it.
Around 80 pages carried the instructions, according to the report.
Meta’s explanation is short: a bug caused it. The company attributed the disclosure to a technical error rather than to any outside attack.
The material involved was the chatbot’s operating guidance, not the personal messages or account details of WhatsApp users.
Why a chatbot’s rulebook is sensitive
A chatbot doesn’t decide its own personality from scratch. Companies give it guidance on tone, topics to steer clear of, and lines it should not cross.
A section titled “WHAT NOT TO SAY” is, by definition, a map of the guardrails. It reveals what a company is worried its product might do if left to its own devices.
Knowing where the guardrails sit can make it easier to probe around them. That is one reason companies tend to treat this kind of material as internal.
What this means for Meta and its AI push
WhatsApp is a particularly sensitive place to put an AI assistant. The app has built much of its identity around private conversations.
Meta has been expanding its AI chatbot capabilities through platforms like WhatsApp and Instagram, including features such as side chats and message assistance tools, with a focus on privacy-enhancing options like “incognito” modes designed to minimize Meta’s access to private conversations.
The company also faces regulatory pressure in Europe concerning its stance on restricting third-party AI chatbots’ access to WhatsApp’s Business API.
A misconfigured web page is not the same as a data breach. Nothing in the reporting suggests users’ messages were compromised.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.