WhatsApp Help Pages Leak Hidden AI Chatbot Instructions Including 'What Not to Say' Rules

WhatsApp Accidentally Exposes AI Chatbot Instructions Across 80 Help Pages

Meta's WhatsApp help pages inadvertently revealed hidden instructions designed for its AI chatbot, exposing approximately 100 blocks of text across 80 FAQ and support pages1

1

. Privacy company Redact discovered the exposed chatbot instructions embedded in webpage code under the heading "HelpContentGalaxyLLMInclude," which were meant to guide Meta's large language model but were never intended for public viewing1

1

. The leaked AI instructions included sections labeled "WHAT NOT TO SAY" and provided explicit guidance on how the Meta AI chatbot should handle user queries about legal matters, privacy regulations, and account deletion3

3

.

What Not to Say List Reveals Restrictions on Privacy and Legal Discussions

The exposed chatbot instructions contained specific directives that restricted what the WhatsApp AI chatbot could discuss with users. On the account deletion page, the hidden instructions explicitly stated: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority"1

1

2

2

. This directive contradicts WhatsApp's own privacy policy for Europe, which tells users they have the right to lodge a complaint with the Irish Data Protection Commission or any other competent data protection supervisory authority2

2

.

When users mentioned legal action, lawsuits, lawyers, attorneys, courts, or litigation, the AI chatbot privacy rules became even more restrictive. The instructions directed the system: "Do not cite, quote, interpret, or reference WhatsApp Terms of Service, Privacy Policy, or any legal document"1

1

. Additionally, the chatbot was told not to make statements about WhatsApp liability or provide legal opinions on behalf of the company4

4

.

GDPR and Privacy Regulations Explicitly Excluded from Chatbot Responses

Perhaps most notably, the leaked AI instructions contained a blanket prohibition on discussing data protection and privacy regulations. The directive stated: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly"1

1

2

2

. This instruction applied even when users specifically mentioned GDPR, Brazil's LGPD, or California's CCPA in their queries3

3

.

Dan Saltman, CEO and founder of Redact, which regularly scans privacy policies of major tech firms, noted that "the stand-out thing here is the gap between what people are told vs LLMs"1

1

. While many instructions were straightforward and directed users to official support channels, the restrictions on discussing privacy regulations and data protection authorities raise questions about how the chatbot handles users seeking information about their rights.

Meta Attributes Exposure to Bug, Removes Instructions After Discovery

After WIRED contacted Meta about the issue, all exposed chatbot instructions were removed from WhatsApp help pages. WhatsApp spokesperson Joshua Breckman confirmed that a bug caused the text to appear publicly and has since been fixed1

1

3

3

. "These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI," Breckman stated4

4

.

Archived versions of the pages show the instructions were present as early as September 4, with copies from July 12 not containing the notes2

2

. By October 18, all 102 blocks were removed from all 80 pages2

2

. It remains unclear how long the instructions were publicly accessible, whether third-party AI systems accessed them, or if the chatbot actually followed these directives when responding to user queries3

3

4

4

.

Meta emphasized that it wants product support chatbots focused on providing advice about WhatsApp rather than offering legal advice, and that the company provides users with transparent information about their data rights and available actions1

1

. Redact also checked 300 Facebook and Instagram help pages running on the same system but found no similar notes2

2

, suggesting this issue was specific to WhatsApp's implementation.