The WhatsApp help pages leak showed the public the hidden instructions behind a big tech support bot. About 80 of WhatsApp’s help and FAQ pages carried hidden text telling Meta’s support chatbot what to avoid, including GDPR, lawyers and data protection regulators.
Meta says a bug caused the leak and that it has removed the text. But the instructions show how a company can quietly shape the answers you get from its AI.
Key takeaways
- About 100 blocks of hidden chatbot instructions sat in the code of roughly 80 WhatsApp help pages.
- Some sections were headed “WHAT NOT TO SAY” and told the bot not to give legal opinions or point users to regulators.
- Privacy company Redact.dev found the text, and WIRED reported the story on 19 October 2026.
- Meta removed the text after WIRED asked about it and blamed a bug.
- Copies can still be found in archived versions of the pages.
What did the WhatsApp help pages leak reveal?
The text was written for a large language model (LLM), the kind of AI that powers chatbots. It didn’t show on screen for people reading the help pages. It sat in the page code under the label “HelpContentGalaxyLLMInclude”, where it served as extra guidance for WhatsApp’s support bot.
Instructions like these are often called a system prompt: the rules a company gives its chatbot before you type a word. It works like the script a manager hands a new call-centre worker. You never see it, but it decides what they will and won’t tell you.
WIRED’s original report reviewed Redact.dev’s findings. It found that many of the instructions were ordinary and mostly sent users to other official support channels. Others were more pointed.
What exactly was the bot told to avoid?
Here are the reported instructions, page by page, and what they mean for you.
| Help page topic | Hidden instruction (as reported) | What it means for you |
|---|---|---|
| Reporting and blocking | If a user mentions lawyers, courts or legal action, don’t cite WhatsApp’s terms or privacy policy and don’t give legal opinions | The bot won’t walk you through the rules that govern your account |
| Privacy questions | Don’t reference data protection laws such as GDPR, LGPD or CCPA, even if the user raises them | Asking about your legal rights gets you a non-answer |
| Account deletion | Don’t recommend contacting any data protection authority or regulator | The bot won’t mention that you can complain to a regulator |
| Accessibility | Treat a user saying they are deaf or hard of hearing as background, not as the support request itself | Mostly about keeping replies on topic |
| New features | Don’t mention community features that aren’t yet generally available | Standard protection against leaking product plans |
GDPR is the European Union’s General Data Protection Regulation. LGPD is Brazil’s version, and CCPA is the California Consumer Privacy Act. All three give you rights over your data. In the EU, those rights include filing a complaint with a data protection authority.
What did Meta say about the leak?
WhatsApp spokesperson Joshua Breckman told WIRED that a bug had made the text visible and that it has been fixed. He said the instructions were meant for the chatbot, not for users. Their job, he said, is to make answers accurate and relevant for people who choose to contact WhatsApp Support with AI.
“We train our support chatbots to be as helpful as possible.”
Joshua Breckman, WhatsApp spokesperson, to WIRED
Meta’s case is that a support bot should give product advice, not legal advice. That’s a fair line to draw, because a chatbot that confidently misreads GDPR could do real harm.
Why does a hidden chatbot script matter?
Dan Saltman, CEO and founder of Redact.dev, summed it up: “The stand-out thing here is the gap between what people are told vs LLMs.” His company says it regularly scans the websites and privacy policies of large tech firms.
In our view, refusing to give legal advice is sensible. Telling a bot never to mention regulators, even on the account deletion page, is a different matter. It doesn’t make answers more accurate. It makes them less complete, and you’d never know.
This matters more as more support starts with a bot. In March 2026, reports said WhatsApp was rolling out an AI assistant as the first reply in its support chat, with a human agent available on request. Small firms are making the same move, as we covered in how AI is transforming small businesses.
Is this a security problem?
Not in the usual sense, because no passwords, messages or user data were reported exposed. Security experts still have a name for this kind of mistake.
OWASP, a nonprofit that publishes widely used security guidance, ranks the top 10 risks for LLM apps. It lists this one as LLM07: System Prompt Leakage. Its advice is blunt: a system prompt “should not be considered a secret, nor should it be used as a security control.” In other words, assume your instructions will be seen one day, and write them so that wouldn’t embarrass you.
What happens next
- Open questions: WIRED said it is unclear how long the text was visible and whether third-party chatbots could read it.
- The archive problem: The text is gone from the live pages but survives in archived copies, so researchers can keep studying it.
- Regulators: As of 20 October 2026, we found no public response from any data protection authority. Still, a rule against pointing users to regulators is the kind of detail regulators tend to notice.
- More scrutiny: Expect privacy researchers to look harder at the hidden instructions behind other companies’ support bots.
FAQ
Were WhatsApp users’ messages exposed?
No. The leak covered instructions written for Meta’s support chatbot. No reports say messages, accounts or personal data were exposed.
Can I still complain to a regulator about WhatsApp?
Yes. Under GDPR, people in the EU can file a complaint with their national data protection authority. The bot’s instructions don’t change your legal rights.
Who found the hidden instructions?
Privacy company Redact.dev found them while scanning big tech websites. WIRED reviewed the text and reported the story on 19 October 2026.
Next time a support bot dodges your question, ask for a human agent and look up your rights on your national data regulator’s website.
1 comment