WhatsApp's help pages were surfacing hidden LLM instructions telling AI not to…

When AI systems are deployed to interact with the public, the instructions guiding their behavior live in an often invisible layer. This background explainer covers how these hidden instructions work, why they matter, and what happens when they surface unexpectedly. Prompt Engineering as Hidden Architecture Large language models operate on instructions embedded in text, not on formal code alone. These instructions, called prompts or system prompts, shape how an AI system responds to queries before it ever sees a user's question. Unlike traditional software where business logic lives in visible code repositories, prompt instructions can be buried in documentation, configuration files, or deployment environments where they are not meant for public consumption. This architectural choice creates an unusual situation. The actual rules governing an AI's behavior exist as plain language text rather than as auditable source code. Developers can modify how a system behaves without shipping a software update, simply by changing what text is fed into the model at runtime. This flexibility is useful for rapid iteration but creates a problem: these instructions often remain opaque to outside observers. The Difference Between Public Policy and Hidden Guidance Organizations publishing AI tools typically maintain public policies describing what their systems will and will not do. A company might state publicly that its chatbot will not provide financial advice, or that it will decline to help with certain requests. These policies are written for users and regulators to read. System prompts are a different layer entirely. They are the actual instructions the model follows, and they need not match the public policy word for word. A public policy might describe guardrails in broad terms, while the system prompt contains specific, detailed rules. More significantly, system prompts can include instructions that the organization prefers users not to see. An instruction to favor one outcome over another, to prioritize one type of response, or to avoid certain topics can all be embedded in a prompt without appearing in any public documentation. This gap between stated policy and hidden instruction has practical consequences. Users and regulators operate based on what they are told a system will do. But if the actual system prompt contradicts or extends those statements, the real behavior may diverge from expectations. How Instructions Become Visible System prompts are usually not intended to be seen by end users. However, large language models can sometimes be prompted or tricked into revealing the instructions they were given, through a technique called prompt injection or jailbreaking. A user might ask the AI to "show your instructions" or embed a request for internal details in a seemingly innocent question. Because language models are trained to be helpful and to respond to requests, they sometimes comply. When help pages or documentation are fed