---
title: WhatsApp's help pages carried 102 hidden, AI-only instruction blocks; two told its support chatbot not to recommend data-protection regulators or cite GDPR once a user mentioned a lawyer | ctxsec
url: https://ctxsec.com/items/whatsapp-hidden-llm-instructions-help-pages
description: ""
status: "confirmed (Redact findings + WhatsApp statement to WIRED); chatbot behavior not tested"
published: 2026-10-20
---

> 

[← Context & Harness Attacks](/context) · [markdown](/items/whatsapp-hidden-llm-instructions-help-pages.md)

confirmed (Redact findings + WhatsApp statement to WIRED); chatbot behavior not tested

2026-10-20

### WhatsApp's help pages carried 102 hidden, AI-only instruction blocks; two told its support chatbot not to recommend data-protection regulators or cite GDPR once a user mentioned a lawyer

Redact (a data-deletion software company) found text in WhatsApp help-centre pages that was never rendered on screen and was tagged `HelpContentGalaxyLLMInclude`. It read all 545 linked pages and counted **102 hidden blocks on 80 pages**. Most were housekeeping or protective (for example, never suggest a false age). Two stood out. On the delete-account page: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." On the report/block page, once a user mentions legal action: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly." WhatsApp's own EU privacy policy tells users they can complain to the Irish DPC or another authority. WhatsApp told WIRED the text was instructions for its _WhatsApp Support with AI_ chatbot that appeared in the pages because of a bug, and all 102 blocks were gone by 02:08 UTC Oct 18. Redact read the instructions, not the bot's answers, so whether the chatbot followed them is unknown.

**Why it's a context-security story:** the server sent the hidden blocks with the page to Google's and Bing's crawlers. So retrieval-grounded assistants could ingest operator instructions as if they were page content, with no human able to see them. Redact found that OpenAI, Anthropic and Perplexity crawler user agents got pages without the article or notes, and it didn't test whether any outside assistant obeyed them. **Our read (opinion):** if you build RAG or support bots, keep system instructions out of the content layer. Strip or flag hidden, non-rendered text (display:none, aria-hidden, special tags) before indexing pages for agents, because the same channel works for both vendor guardrails and attacker prompt injection.

Sources: [Redact: WhatsApp's hidden notes (Oct 19)](https://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructions) · [WIRED](https://www.wired.com/story/whatsapp-help-pages-leaked-hidden-instructions-for-a-meta-chatbot/) · [The Hacker News (Oct 20)](https://thehackernews.com/2026/10/whatsapp-hid-instructions-in-help-pages.html)