Industry

Meta's Hidden Chatbot Instructions Exposed in WhatsApp Help Pages

Privacy researchers discovered around 80 WhatsApp support pages accidentally revealed internal LLM prompts instructing the chatbot to avoid legal advice and regulatory referrals.

Last verified:

Meta’s Leaked Chatbot Instructions Reveal Guardrails on User Support

Privacy research firm Redact discovered that approximately 80 WhatsApp help pages inadvertently exposed internal instructions written for Meta’s customer-support chatbot, according to reporting by Wired AI. The hidden text, containing around 100 instruction blocks, was embedded in the HTML code under labels such as “HelpContentGalaxyLLMInclude” and included specific guidance on topics Meta wanted the chatbot to avoid—most notably legal advice and referrals to regulatory authorities.

What the Hidden Instructions Revealed

According to Wired’s investigation, the exposed prompts included directives such as “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” when users asked about account deletion. On pages addressing reporting and blocking, the instructions instructed the chatbot not to cite WhatsApp’s Terms of Service, Privacy Policy, or make statements about company liability. The guidelines also explicitly told the model to refrain from offering legal opinions or interpretations on behalf of WhatsApp.

Redact CEO and founder Dan Saltman told Wired that while many of the exposed instructions were straightforward—often directing users to official support channels—they were clearly not intended for public visibility. The instructions appeared to function as a safety layer to prevent the chatbot from making potentially problematic statements about legal matters or regulatory compliance.

Meta’s Response and Patch Timeline

WhatsApp spokesperson Joshua Breckman confirmed to Wired that the exposure resulted from a technical bug and stated the text had been removed across all pages following the publication’s inquiry. Breckman characterized the instructions as standard guidance designed to “ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI,” emphasizing that the prompts were directed at the AI system, not users. The company declined to specify how long the pages had exposed the text or whether third-party systems accessed it.

Why This Matters

The incident underscores how template-based chatbot guardrails—necessary for risk management in customer-facing AI—can become liabilities if exposed, revealing both operational constraints and potentially sensitive policy positions. For organizations deploying support chatbots, this highlights the need for strict separation between public-facing content and embedded system instructions, particularly when those instructions reference legal or regulatory limitations. The discovery also illustrates how privacy researchers are increasingly scanning major tech platforms’ code for unintended data leakage, creating a new category of disclosure risk for companies deploying LLM-based services at scale.

Frequently Asked Questions

What exactly was leaked in the WhatsApp help pages?

Around 100 blocks of hidden text containing system instructions for Meta's internal chatbot, embedded in the HTML code of approximately 80 support pages. These instructions were not visible to regular users but appeared in the page source code.

How did the instructions end up on public pages?

According to WhatsApp spokesperson Joshua Breckman, the exposure was caused by a bug in how the chatbot instructions were deployed to the web pages. Meta removed the text after Wired contacted the company about the discovery.

Could third-party chatbots access these instructions?

Wired's reporting does not clarify whether external AI systems could access the exposed text. Meta has not provided details on the scope of potential access during the exposure period.

#meta #whatsapp #ai-safety #prompt-injection #privacy #chatbots