WhatsApp Help Pages Exposed Hidden Meta Chatbot Instructions
A technical bug revealed internal system prompts intended for Meta's support AI.
1System Prompt Exposure
Dozens of WhatsApp help pages inadvertently exposed internal instructions designed for Meta's large language model. These hidden text blocks, labeled HelpContentGalaxyLLMInclude, were visible within the underlying source code of approximately 80 FAQ pages. The leak provided a rare look at the specific directives governing how the company's support chatbot interacts with users.
2Legal Advice Restrictions
The exposed instructions explicitly forbade the chatbot from providing legal advice or interpreting documents like the Privacy Policy. The system was programmed to avoid referencing data protection regulations such as GDPR, LGPD, or CCPA. Even when users explicitly mentioned lawsuits or attorneys, the AI was directed to refrain from making statements regarding liability.
3Regulatory Avoidance Tactics
One specific instruction directed the chatbot to avoid recommending that users contact Data Protection Authorities or any supervisory regulators. This directive suggests a strategic effort to keep support interactions focused on product functionality rather than legal escalation. The instructions aimed to prevent the AI from acting as a legal representative for the company.
4Operational Chatbot Constraints
Beyond legal boundaries, the chatbot received instructions on handling specific user demographics and feature disclosures. For instance, the AI was told to treat mentions of hearing impairment as background information rather than a support request. Additionally, the system was strictly prohibited from referencing upcoming community features that were not yet generally available to users.
5Discovery and Analysis
The exposure was identified by Redact, a privacy company that regularly scans the websites and policies of major technology firms. CEO Dan Saltman noted a significant gap between the information presented to human users and the internal instructions governing the LLM. The findings highlight the risks associated with embedding system prompts directly into web code.
6Meta's Official Response
Following inquiries from WIRED, Meta removed the hidden text from all affected pages. Spokesperson Joshua Breckman attributed the incident to a technical bug that has since been resolved. He emphasized that these instructions were intended solely for the chatbot to ensure accurate responses and were never meant to be publicly accessible to users.
7Technical Implementation Risks
The incident underscores the challenges of managing AI-driven support systems within existing web infrastructure. By embedding LLM instructions directly into the HTML of help pages, the company created a vulnerability where sensitive operational logic could be viewed by anyone inspecting the page source. This highlights the need for more secure prompt management architectures.
8Implications for AI Transparency
While the company maintains that its goal is to provide helpful support, the leaked instructions reveal the complex guardrails required to manage automated customer service. The incident raises questions about how tech firms balance AI efficiency with transparency. Moving forward, companies must ensure that system-level directives remain isolated from public-facing web content.
Original article: Wired — WhatsApp Help Pages Exposed Hidden Meta Chatbot Instructions