WhatsApp Leaked Its Own Chatbot Instructions
Original: WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot
Why This Matters
Leaked system prompts reveal how platforms quietly shape AI responses on sensitive legal topics.
Around 80 WhatsApp help pages inadvertently exposed hidden LLM instructions in their source code, including sections titled 'WHAT NOT TO SAY' and directives to avoid legal or regulatory topics. Meta confirmed a bug caused the leak.
Privacy company Redact.dev found roughly 80 WhatsApp FAQ and help pages with hidden text blocks—labeled 'HelpContentGalaxyLLMInclude'—embedded in the page source but invisible to ordinary visitors. The ~100 blocks of instructions were designed to guide one of Meta's LLMs powering WhatsApp Support. WIRED independently verified the exposed text, which remains visible in archived page versions.
Some of the instructions were routine, directing the bot toward official support channels. Others were more striking. On the page covering account deletion, the chatbot was told: 'Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.' On a page about reporting and blocking, the LLM was instructed not to cite WhatsApp's Terms of Service or Privacy Policy, not to interpret GDPR, CCPA, or LGPD even when a user explicitly raises them, and not to provide 'legal opinions, defenses, or interpretations on behalf of WhatsApp.'
After WIRED contacted Meta, all instructions were stripped from the live pages. WhatsApp spokesperson Joshua Breckman attributed the exposure to a bug and said the instructions 'are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI.' It remains unclear how long the text was exposed or whether external crawlers or third-party models picked it up.