By Hamza Faisal

Dozens of WhatsApp's official help pages were quietly carrying hidden instructions for a Meta AI chatbot inside their code — including directions telling the bot not to suggest users contact data protection regulators. The text was never visible to visitors, but privacy researchers found it buried in the pages' HTML, offering a rare unscripted look at how Meta coaches its AI support assistant.

The discovery comes from Redact, a privacy company that routinely scans big tech sites and privacy policies. The team examined 545 WhatsApp help pages and found roughly 100 blocks of hidden text addressed to an AI system across about 80 of them — text written not for humans, but to guide how Meta's own support chatbot answers questions.

What the hidden notes said

Some of the instructions were filed under headings bluntly labeled "WHAT NOT TO SAY." On the page explaining how to delete your account, the chatbot was told: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority" about the issue. Archived copies confirm that note was live by early September.

Other pages were just as careful around anything legal. On the help page about reporting and blocking users, the guidance said that if a user mentions legal action or lawyers, the bot should not cite or interpret WhatsApp's Terms of Service or Privacy Policy — and should not make statements about the company's liability.

The caution extended to privacy law itself: the chatbot was told not to reference laws like the GDPR, Brazil's LGPD, or California's CCPA — even when users raised them directly. Other notes covered instructions for deaf or hard-of-hearing users and barred the bot from mentioning unreleased features.

How the leak was found — and what Meta says

Redact's researchers found the notes under a code heading called "HelpContentGalaxyLLMInclude." WIRED independently viewed the exposed text in archived versions of the pages and contacted Meta — after which the instructions disappeared from every page.

A WhatsApp spokesperson, Joshua Breckman, told WIRED the exposure was caused by a bug that has now been fixed. "These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses," Breckman said, adding that the company wants its support chatbots focused on helping with WhatsApp issues rather than dispensing legal advice.

The researchers are careful about what the evidence does not prove: they read the chatbot's instructions, not its actual answers, so it is unclear whether the bot ever followed the notes or how long they were exposed.

Why it matters

"The stand-out thing here is the gap between what people are told vs LLMs," says Dan Saltman, Redact's CEO and founder. The find pulls back the curtain on AI prompt engineering — the behind-the-scenes process by which companies teach chatbots to dodge risky topics and steer users away from paths that could lead to legal or regulatory trouble.

As Redact's authors put it: "Naming a law that protects you isn't legal advice, and neither is telling you a regulator exists."

By Hamza Faisal