“Don’t Mention the Regulator”: WhatsApp's Hidden AI Instructions Get Exposed

Your help pages have two texts: one for you, one for the AI. On October 19, privacy company Redact published research with an uncomfortable finding: of 545 pages in WhatsApp's help center, 80 contained 102 hidden text blocks in their source code — instructions written for AI to read, invisible to humans. Two of those blocks are chilling.

The first sat on the "how to delete your account" page. What users saw: accounts are deleted 30 days after deactivation or after 120 days of inactivity (365 in the US). What the AI received: the same waiting periods, plus three orders — chief among them "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority."

The second sat on the "reporting and blocking" page. The moment a user mentioned legal action, lawsuits, lawyers, or courts, the AI was told to stop citing or interpreting any privacy law — including GDPR, LGPD, or CCPA the user had raised themselves. In Redact's words: "WhatsApp's privacy policy for Europe tells you that you can complain to a data protection regulator, and it names the law that gives you that right. Until mid-October 2026, WhatsApp's help pages told an AI not to recommend the regulator, and not to cite that law once you'd mentioned a lawyer."

The sharpest irony: WhatsApp's European privacy policy states in black and white that users may complain to the Irish Data Protection Commission "or any other competent data protection supervisory authority" — the policy humans read says "you can go to the regulator"; the instructions the AI received said "don't tell them."

102 hidden blocks: where they hid, how they were found

The instructions lived in page source inside a code block named HelpContentGalaxyLLMInclude — LLM being short for large language model. Invisible on screen, readable by anything crawling the page. Redact first spotted them on October 8, 2026, then read all 545 pages the help center links to, archiving each one: 80 pages with hidden AI text, 102 blocks total.

Redact read every block: most were harmless housekeeping — disambiguating confused terms, pointing questions to better pages, describing paid plans at length; a few protected users (like compromised-account guidance); a handful were milder cousins of the two above. The report ends with summaries of the 20 most interesting blocks.

WIRED independently verified the research: reporters viewed the exposed text in archived page source themselves, then contacted Meta — after which the instructions vanished from every WhatsApp page. WhatsApp spokesperson Joshua Breckman called it a bug, now fixed: "These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses."

"Naming a law that protects you isn't legal advice"

One sentence in Redact's report deserves full quotation: "Naming a law that protects you isn't legal advice, and neither is telling you a regulator exists."

It draws the crucial line. A company not wanting its support chatbot to dispense legal opinions in its name is entirely reasonable — telling people to see a lawyer is the right call. But "don't give legal advice" and "hide public rights information" are different things. Users' rights under GDPR Article 17 (erasure) and Article 77 (complaint) don't evaporate because of a chatbot instruction; what the instruction could do was make sure you never heard about those rights in the place you went for help.

Redact founder Dan Saltman's comment to WIRED cuts deeper: "The stand-out thing here is the gap between what people are told vs LLMs." — system prompts are becoming a new, opaque governance layer: they decide what an AI may and may not say, and users never see them.

The crawlers read it too: your system prompt is a liability

The report holds one more detail that matters most to developers: WhatsApp's servers sent these hidden instructions, along with the rest of the page, to Google's and Bing's crawlers. When Redact requested pages on October 9 under the crawler names used by OpenAI, Anthropic, and Perplexity, they came back without the article or the notes — but Google and Bing got the full version.

Plainly: system prompts hidden in your HTML source get read, archived, and indexed by search crawlers. Today it's your support chatbot's instructions; tomorrow it could be your pricing strategy, internal talking points, or unreleased feature descriptions. The warnings for vibe developers are direct:

• Don't put system prompts in frontend code: anything "hidden" in HTML/JS source is public to crawlers and to anyone who hits View Source. Prompts belong server-side only.
• When writing a chatbot's "don't say" list, ask one question: if this instruction made tomorrow's headlines, could I defend it? "Don't give legal advice" — defensible. "Don't tell users the regulator exists" — not.
• Audit your own site through a crawler's eyes, regularly: fetch your pages with Google's crawler user agent and see what the AI sees — your robots.txt and rendering strategy define you in the AI's eyes.
• Have a lawyer review the support chatbot's "legal red lines": not "legal is too slow so we skipped it" — the red lines themselves are a legal product.

A bug, or a systemic problem?

WhatsApp calls it a bug. But several facts make the "bug" explanation thin: first, 102 blocks across 80 pages spanning account deletion, reporting, bans, and payments — "bugs" are rarely this organized; second, Facebook's and Instagram's help centers run on the same system, and Redact checked 300 of their pages — not one had similar hidden instructions — so this wasn't default system behavior, someone wrote these for WhatsApp specifically; third, archives show the notes present at least since September 4 (absent in the July 12 copy) — they ran quietly for over a month.

In fairness, Redact stated its own evidentiary limits: they read the instructions, not the chatbot's actual answers — they cannot confirm the bot ever followed the notes or for how long. That's the restraint honest researchers owe, and this article reports it as stated.

Bug or not, the incident tears open a seam: in 2026, every company's website source may be carrying a set of "hidden text for AI" — the SEO era hid keywords for search engines; the AI era hides instructions for chatbots. Redact says it "regularly scans big tech's sites and privacy policies" — WhatsApp is just the first to get exposed. Who's next is worth every AI-support team auditing before someone else does it for them.

Transparency note

Core facts (545 pages / 80 pages / 102 blocks, first spotted October 8, 2026, the HelpContentGalaxyLLMInclude block name, Google/Bing crawlers receiving the full version, zero similar notes in 300 Facebook/Instagram pages) are from Redact.dev's original October 19, 2026 research report (page-stamped 2026-10-19T16:00:00+00:00), which includes full reproduction steps (Internet Archive copies + view source + search "Do not recommend contacting"). WIRED's independent verification and spokesperson Joshua Breckman's response are cross-confirmed via multiple outlets. Redact explicitly states it did not verify whether the chatbot acted on the notes or exactly how long they were present (archives show at least September 4, 2026 to mid-October) — reported here as stated.