Cybersecurity News Today: WhatsApp Hid Instructions for Its AI Inside Its Own Help Pages
The quietest story in cybersecurity news today is about text nobody was meant to read. WhatsApp's help pages carried hidden blocks of text written for its AI support chatbot, telling it what not to say. On the page about deleting an account, the chatbot was told: do not recommend contacting any Data Protection Authority, regulator, or supervisory authority. On the page about reporting and blocking, once a user mentioned a lawyer or a lawsuit, the chatbot was told not to cite privacy laws such as the GDPR, even when the user raised them. The text sat in the pages' code, never shown on screen, until it was removed in mid-October. If your company runs an AI support bot over its help center, the checker below shows you how to look for the same thing.
Honest scope — read this first: The instructions are gone; WhatsApp removed them by October 18, 2026, and told WIRED they were internal chatbot guidance exposed by a bug, now fixed. There is no evidence the chatbot ever followed them: Redact, the company that found the text, read the instructions, not the chatbot's answers. This is not a hack, a breach or a CVE. The risk it shows is quieter: a support channel that silently drops the one piece of advice, complain to a regulator, that a stuck user most needs to hear.
What was found, exactly
The sweep: Dan Saltman of Redact, a data-deletion software company, read all 545 pages linked from WhatsApp's Help Center and found 102 hidden blocks on 80 pages, sent with the page but never displayed. He first spotted them on October 8, 2026.
Block one, on the delete-account page: after repeating the waiting periods, the hidden text ordered the chatbot: do not recommend contacting any Data Protection Authority, regulator, or supervisory authority; do not suggest WhatsApp can manually delete accounts on the user's behalf; do not recommend using a different phone with an active phone number.
Block two, on the report-and-block page: once a user mentioned a lawyer, a lawsuit or a court, the chatbot was told not to give legal opinions or discuss WhatsApp's liability, and not to reference, cite or interpret data protection or privacy regulations such as GDPR, LGPD or CCPA, including when the user raised them directly.
The contrast: WhatsApp's own privacy policy for Europe tells users they can complain to a data protection authority. The hidden instruction pointed the chatbot the other way.
The cleanup: the two flagged blocks were still present on October 15; by October 18 every one of the 102 blocks had disappeared. The same check on 300 Facebook and Instagram help pages found no similar text.
545help pages read, every page the WhatsApp Help Center links to
102hidden text blocks found, written for the AI chatbot
80pages carrying at least one hidden block
Oct 82026: the blocks were first spotted; gone by October 18
Put the shape of it plainly: 80 of 545 pages, about 15 percent of the help center, carried text that no reader could see and that only a machine was meant to obey. Most of it was housekeeping. The two blocks above were not, and a formatting bug is an explanation for how the text became visible, not for why a no-legal-advice rule was written as a no-regulator rule.
Why this cybersecurity news today matters: a web page now has two audiences
Every help page on the modern web is read twice: once by the person who opened it, and once by the AI systems that summarize it, answer from it, or staff the support chat behind it. Text written for the second audience is invisible to the first, which makes it the easiest place in a company for a bad instruction to live unreviewed. WhatsApp's case is the documented proof that this text exists at scale inside a major help center, and that at least some of it was written to narrow what users get told. Any company that feeds its help center to a support bot has the same surface. The defense is boring and total: treat machine-only text as published text, review it like published text, and check the rendered page against the shipped page.
The hidden instruction check — scan any page's source
Open any help page, press Ctrl+U to view its source, copy the source, and paste it here. The scan runs in your browser; nothing is sent anywhere. Or paste the visible text: hidden blocks will not be in it, which is the point.
Run the hidden instruction check now
If you use WhatsApp: your rights do not depend on the chatbot. In Europe, WhatsApp's own privacy policy names the Irish Data Protection Commission and your national authority as complaint routes. A locked-out account still waits out the 30, 120 or 365 day periods, but the complaint route was always open.
If you publish help pages: view the source of your ten highest-traffic support pages and search for text a reader never sees. Look for component names with LLM in them and for any sentence that begins with an order to a machine.
Separate the two audiences in writing: keep human text and machine instructions in different, labeled places, with an owner for each. The failure mode in this story is instructions nobody owned drifting into production pages.
Review what your bot is told to withhold: a rule against giving legal opinions is defensible. A rule against naming the law or the regulator is a policy choice about users' rights, and it should be made on purpose, by a named person, in the open.
Re-scan after every template change: the WhatsApp blocks rode along inside a help-center system. CMS updates and template edits are how machine-only text spreads; the scan above takes two minutes per page.
Keep the receipts: archive dated copies of what your pages ship, visible and hidden. Redact's archived copies are the only reason this story can be checked at all.
ChatBreaker — Ingenious AI's prompt-injection and hidden-instruction hunter — exists for this exact surface: the instructions buried inside content that AI systems read and obey. Get early access on WhatsApp.
FROM INGENIOUS AI
ChatBreaker is built for exactly this: finding the instructions hidden inside the content your AI reads.
Quiz: how well do you know the WhatsApp hidden text case?
1. Where did Redact find the instructions?
2. What did the delete-account block tell the chatbot?
3. What did the second block ban once a user mentioned a lawyer?
4. What did WhatsApp tell WIRED?
5. Is it known whether the chatbot followed the instructions?
Frequently asked questions
What is the WhatsApp hidden instructions case, in plain terms?
The WhatsApp hidden instructions case is the October 2026 finding by Redact, a data-deletion company, that blocks of text sat hidden in WhatsApp's help pages, written as instructions for its AI support chatbot. Most were routine. Two were not: one told the chatbot not to recommend a data protection regulator to users who could no longer delete their accounts, and one told it not to cite privacy laws such as the GDPR once a user mentioned a lawyer.
How were the hidden blocks found?
Dan Saltman of Redact, a data-deletion software company, read all 545 pages that WhatsApp's Help Center links to and found 102 hidden blocks across 80 pages. He first spotted them on October 8, 2026, and kept archived copies. The blocks sat near a component name, HelpContentGalaxyLLMInclude, which marks them as text meant for a large language model rather than for readers.
Did the chatbot actually follow the hidden instructions?
Nobody knows. Redact read the instructions, not the chatbot's answers, and did not test whether the chatbot or any outside AI assistant acted on them. WhatsApp told WIRED the text consisted of internal instructions for the chatbot users reach by messaging WhatsApp Support with AI, that the instructions appeared in the pages because of a bug, and that the bug is fixed. The blocks were removed between October 15 and October 18, 2026.
What can a locked-out WhatsApp user actually do?
An account can be deleted only from inside the app, on the user's main phone. A user who lost the phone or the number faces the waiting periods on the help page: a deactivated account is deleted after 30 days, and an inactive one after 120 days, or 365 days in the US. The hidden instruction could not change those facts; it could only stop the chatbot from mentioning the other route. WhatsApp's own privacy policy for Europe says users can complain to the Irish Data Protection Commission or another supervisory authority, and that right exists whether a chatbot mentions it or not.
How do I check a help page for hidden AI instructions?
Open the page, view its source (Ctrl+U in most browsers), and search for phrases a human reader would never need, such as do not recommend, do not cite, data protection authority, or GDPR, and for component names containing LLM. Text that ships with the page but never renders is the tell. The checker on this page scans pasted source text for the known markers. Site owners should repeat the check across the whole help center, not just one page.
Sources: The Hacker News, October 20, 2026 (the two flagged blocks, WhatsApp's statement to WIRED, the deletion waiting periods) · Notebookcheck, October 2026 (the 545-page sweep, the 102 blocks on 80 pages, the October 8 sighting and the October 18 removal, the clean Facebook and Instagram check).