Dozens of customer support pages on WhatsApp’s official website inadvertently exposed internal system prompts and operational rules intended for an artificial intelligence chatbot, according to reporting by WIRED. The exposed materials included proprietary guidance detailing rules such as "WHAT NOT TO SAY" to users.
Parent company Meta attributed the public exposure of the internal AI instructions to a technical bug, which permitted backend system details to appear across approximately 80 individual pages on the WhatsApp website.
Exposure of Proprietary System Prompts
The leaked information offered an unvarnished look into the exact operational boundaries configured for the automated assistant. Behind typical customer-facing interfaces, AI systems rely on structured system prompts that establish conversational boundaries, acceptable responses, and topics that automated agents must strictly avoid.
On the affected WhatsApp pages, users navigating the website could view the explicit constraints programmed into the tool, including directives cataloged under "WHAT NOT TO SAY." Such guidelines are standard engineering components used to prevent automated software from generating inappropriate answers, offering unreliable advice, or breaching company safety protocols.
Meta Points to Technical Flaw
Following the discovery, Meta confirmed that the public visibility of the internal chatbot instructions stemmed from an unexpected software flaw rather than an intentional change in website documentation. The company stated that a bug caused the information to surface across the roughly 80 online pages.
The incident highlights ongoing technical hurdles for large technology companies as they integrate generative AI systems across their existing web ecosystems. Backend administrative rules and guardrails, which are typically hidden from public view, remain vulnerable to inadvertent exposure if content delivery systems fail to isolate automated prompts from outward-facing reference pages.
