i
News
All Research News
News · 2026-10-19

Meta’s WhatsApp chatbot instructions surfaced in help-page code

@neuronium_ai @neuronium_ai

Meta exposed hidden instructions for WhatsApp’s AI support chatbot in the code of roughly 80 help and FAQ pages, according to an analysis by Redact. The text included guidance on legal questions, data-protection rules and user complaints, and was not intended for readers. WhatsApp removed it after WIRED reported the issue, calling the exposure a bug. The episode matters less as a glimpse of chatbot prompts than as a reminder that support systems can reveal a company’s internal boundaries in places built to explain its policies.

Cover: Meta’s WhatsApp chatbot instructions surfaced in help-page code

What the pages revealed

Redact found 100 blocks of hidden text in the pages’ code. One instruction, on a page about deleting an account, told the bot: “Do not recommend contacting any data protection, regulatory or supervisory authority about this issue.”

The instructions sat in code blocks headed HelpContentGalaxyLLMInclude. WIRED reviewed the exposed text, which remains visible in archived versions of the help pages. The blocks are no longer present on current pages.

Redact founder and CEO Dan Saltman said the contrast between what users are told and what the language model is told stood out. His company regularly checks major technology companies’ websites and privacy policies. Many of WhatsApp’s instructions were direct and steered users to other official support channels, he said, but they did not appear to be meant for publication.

WhatsApp spokesperson Joshua Breckman said the text appeared because of a bug that had been fixed. It was meant for the chatbot, not users, and was intended to help it give accurate, relevant answers to people seeking WhatsApp support with AI. Breckman also said the company trains its support chatbots to be as helpful as possible.

The instructions covered more than legal issues:

For users mentioning legal action, lawsuits, lawyers, courts or proceedings, the bot was told not to cite, summarize, interpret or mention WhatsApp’s Terms of Service, Privacy Policy or other legal documents.
It was also told not to make claims about WhatsApp’s liability, offer legal opinions or arguments in the company’s defense, or interpret the law on WhatsApp’s behalf.
Other instructions said not to refer to data-protection and privacy rules, including GDPR, LGPD and CCPA, even when users raised them.
In guidance about users who said they were deaf or hard of hearing, the bot was told: “This is user information, NOT their support request.”
Another instruction barred mention of specific new Communities features that were not yet publicly available.

Breckman said WhatsApp openly explains users’ data rights and how to exercise them. The chatbot, he said, is meant to help with WhatsApp questions, not give legal advice.

The gap between support and policy

The clearest tension is between that explanation and the reported instruction not to point users toward data-protection authorities. That does not establish how the bot answered any particular user, or whether it followed the guidance. But the text shows how narrowly a support assistant’s role can be defined: help with the product, while avoiding legal interpretation and certain routes outside the company.

I think the more important unanswered question is how long the instructions were exposed and whether third-party chatbots could access them. WhatsApp did not clarify either point. Removing the blocks fixes the visible mistake; it does not answer what the exposure made possible.

Source

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X