WhatsApp help pages leak orders telling its AI not to mention regulators

Around 80 of WhatsApp's help and FAQ pages carried 102 blocks of text you never saw. They weren't drawn on screen, but they sat in the page code under a label called HelpContentGalaxyLLMInclude, and they were written for WhatsApp's own support chatbot. A privacy company spotted them on 8 October, then read all 545 pages the help centre links to.

The two that stood out

Most of the blocks were housekeeping: clearing up confusing terms, pointing questions to better pages, a few even protecting users. Two weren't. On the delete-account page, for people who can't delete their account because they've lost their phone or number, the bot was told: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.”

On the reporting and blocking page, once a user mentioned legal action, lawyers or courts, the bot was told: “Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly.” That sits awkwardly next to WhatsApp's own European privacy policy, which, the researchers point out, names the Irish Data Protection Commission and the GDPR when it explains your right to complain.

How it leaked

Archived copies suggest the notes appeared sometime between 12 July and 4 September. WhatsApp's server sent them along with the rest of each page, including to search engine crawlers. Nobody has shown whether the chatbot actually followed them, or whether any outside AI assistant picked them up.

What WhatsApp said

By 18 October all 102 blocks were gone from all 80 pages. A WhatsApp spokesperson said: “These are instructions to our chatbot, not directly to users,” and that they'd been shown because of a bug that's now fixed. The company says it wants its support chatbots “focused on providing the best advice to users about WhatsApp” and “not to offer legal advice.”

Our take

Keeping a support bot out of legal advice is fair enough. Telling someone a regulator exists isn't legal advice, though. Two lessons for anyone shipping a bot. First, anything you send to the browser is public, including prompts tucked into hidden elements. If your bot's instructions would embarrass you in a screenshot, assume they'll end up in one. Second, a bot's instructions are policy. If they quietly contradict your published privacy policy, people will believe the hidden version.