Imagine you are trying to figure out how to delete your WhatsApp account, or perhaps you are curious about what happens when you report a harassing message. You land on one of the company’s official help pages, expecting a straightforward, human-written explanation. But hidden beneath the visible text, buried in the webpage’s background code, there is another layer of messaging entirely — one meant not for you, but for Meta’s artificial intelligence. A recent investigation by the privacy firm Redact has revealed that dozens of WhatsApp web pages contained secret instructions embedded in their code, designed to guide the company’s own chatbot in how to respond to users. These were not just technical commands; they were explicit, sometimes surprising directives that told the AI what it should and, perhaps more importantly, should not say. Around eighty help and FAQ pages were found to contain these hidden blocks, with roughly a hundred separate pieces of instruction. Some of them were bluntly labeled with headings like “WHAT NOT TO SAY,” offering a rare, unintended glimpse into the carefully guarded world of AI prompt engineering — the behind-the-scenes process by which companies teach their chatbots to behave, avoid risky topics, and stay on message.
The content of these hidden instructions reveals a deep institutional caution, particularly around anything that touches the law. For instance, on the help page about reporting and blocking users, the embedded guidance told the chatbot that if a user mentions legal action, lawsuits, lawyers, courts, or litigation, it should not engage with the subject at all. The instructions were very specific: do not cite, quote, interpret, or reference WhatsApp’s Terms of Service, its Privacy Policy, or any legal document. The AI was also told not to make any statements about WhatsApp’s liability or non-liability, and it was explicitly instructed not to provide legal opinions, defenses, or interpretations on behalf of the company. Even more striking, the instructions extended to broader privacy regulations, telling the chatbot not to reference, cite, or interpret laws such as the GDPR in Europe, Brazil’s LGPD, or California’s CCPA — even when users raised those laws directly. Another page, this one about deleting your account, contained a directive that was almost startling in its bluntness: do not recommend contacting any Data Protection Authority, regulator, or supervisory authority about the issue. That specific instruction was discovered in the code of an archived version of the page, and it paints a picture of a company working hard to keep its AI support assistant within narrow boundaries, steering users away from any path that might lead to legal entanglements or external oversight.
When WIRED contacted Meta about the discovery, the company’s response was quick. The exposed text was removed from across all the affected pages, and a WhatsApp spokesperson, Joshua Breckman, explained that the text had been displayed due to a bug that has now been fixed. He clarified that these were instructions intended for the company’s own chatbot, not direct messages to users, and that they were meant to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI. In other words, the hidden blocks were part of a prompt template — a set of guardrails for an AI system that has likely become the first point of contact for many users seeking help. It remains unclear exactly how long the text had been exposed in the page code, whether it could have been accessed or scraped by third-party chatbots, or why the bug happened in the first place. Breckman defended the underlying approach, saying the company wants its support chatbots to be focused on providing the best advice to users about WhatsApp itself, not to offer legal advice. He also stressed that WhatsApp provides users with transparent information about their data rights and how they can take action. But the incident raises uncomfortable questions about how much of our digital experience is mediated by AI, and how often the rules governing that AI are hidden from the very people it is meant to serve.
The discovery itself was something of a routine scan turned into a revelation. Dan Saltman, the CEO and founder of Redact.dev, explained that his company regularly scans the websites and privacy policies of large technology firms, looking for exactly this kind of hidden data. He noted that many of the WhatsApp instructions the team found were actually straightforward and pointed users toward official support channels, which is not inherently sinister. But the standout feature, in his words, was the gap between what people are told and what the large language model is instructed to do. That gap is worth pausing over. On the surface, WhatsApp’s help pages project a friendly, informative tone, offering advice on privacy, reporting, and account management. But underneath, the AI is being carefully scripted, with explicit warnings about what it can and cannot say. It is not unusual for companies to train support chatbots to avoid legal liability, of course. Any customer service representative, human or digital, is typically trained not to give legal advice or make promises the company cannot keep. What is unusual is that the general public got to see the script. The hidden directions were embedded under a code heading that read “HelpContentGalaxyLLMInclude,” suggesting they were part of a larger system for injecting AI instructions into help content. That technical detail may sound mundane, but it reveals how deeply integrated AI has become in the most ordinary corners of the web — even the page that tells you how to block a contact is now powered by a hidden layer of machine-readable command.
Beyond the legal cautions, the exposed instructions included a few other curious details that show just how carefully Meta is managing its chatbot’s responses. For example, one set of instructions dealt with users who mention that they are deaf or hard of hearing. The AI was told to recognize that such a statement is background information about the user, not their actual support request. This seems like an attempt to prevent the chatbot from misunderstanding a user’s personal context as the main issue, perhaps to avoid inappropriate tangents or to ensure the AI focuses on solving the concrete problem at hand. Another instruction warned the chatbot not to mention specific upcoming community features that were not yet generally available. That kind of directive is understandable from a product management perspective — companies do not want chatbots leaking unannounced features or promising things that are still in development — but it also illustrates the degree of control that developers are trying to exert over conversational AI. Every turn the chatbot takes, every piece of information it shares, is being filtered through layers of caution and corporate strategy. These instructions were not necessarily malicious. In fact, many of them sound like reasonable operational guidelines for a customer service bot. But they were never meant to be seen, and their accidental exposure gives us a rare, unfiltered look at how a major tech company thinks about its AI: as a powerful but risky tool that must be carefully hemmed in with rules, exceptions, and explicit prohibitions.
What does all of this mean for the average WhatsApp user? On one level, it is a reminder that the help pages we take for granted are no longer simply static documents written by humans. They are increasingly dynamic interfaces for AI systems, and behind those interfaces lies a complex structure of prompts, safeguards, and hidden instructions that shape the conversation before it even begins. The fact that a user reading a help page might not see the chatbot instructions is perhaps not a surprise, but the idea that those instructions include explicit prohibitions on telling users about data protection authorities or legal rights is unsettling. It suggests that the primary goal of the support chatbot is not to empower users with every possible avenue of recourse, but to keep the conversation safe, contained, and within the walls of WhatsApp’s own ecosystem. There is a human story here too. The people who ask for help on WhatsApp are often dealing with real problems: a hacked account, a harassing message, a lost chat history, a fear that their private data has been misused. They come to the help pages looking for clarity and reassurance. Instead, they may be talking to an invisible AI whose script has been written to avoid certain topics entirely. That is not necessarily deception — Meta argues it is simply a way to keep support focused and avoid giving legal advice — but transparency matters. The accidental exposure of these instructions shows that the machinery of AI support is still messy, still prone to bugs, and still evolving in ways that the public barely understands. Perhaps the most humanizing takeaway is this: even the most powerful companies in the world are still figuring out how to talk to us with machines, and sometimes their hidden scripts slip out, reminding us that behind every chatbot is a team of people making hard choices about what to say, what to hide, and what to leave out entirely.
