WhatsApp's public privacy policy recognizes users' right to complain to data protection regulators. Still, hidden instructions on its Help Center pages told its AI chatbot not to recommend that option in certain cases.
Data deletion company Redact uncovered the instructions embedded in the HTML source code of WhatsApp's Help Center pages. Some directed the chatbot away from recommending regulatory complaints over account deletion problems, while others told it not to cite certain information.
WhatsApp has attributed the exposure to a bug and has since removed the instructions. The discovery does not establish that WhatsApp violated privacy law. However, it highlights the need to ensure AI responses align with a company's published policies.
What’s the issue at the center of this story?
Redact found 102 blocks of hidden AI instructions across 80 WhatsApp Help Center pages.
The instructions were embedded in the pages' HTML source code, making them invisible to ordinary visitors but potentially available to systems processing the underlying page content. Redact found instructions under the identifier HelpContentGalaxyLLMInclude, intended to mark content for WhatsApp’s AI support chatbot.
In one cited example, WhatsApp explained what users could do if they lost access to their phone or number. However, the hidden instructions went further, telling the chatbot not to recommend contacting a data protection authority or suggest that WhatsApp could manually delete the account.
That creates a disparity between the options recognized in the company’s public privacy information and the guidance its chatbot was instructed to provide.
Redact identified 102 blocks of AI-directed instructions across 80 WhatsApp Help Center pages. Not all restricted information: some were intended to clarify confusing topics, direct users to more relevant pages, or provide safety guidance.
WhatsApp says a bug exposed the instructions
WhatsApp confirmed to WIRED that the instructions appeared on its Help Center pages because of a bug. The company also said its AI support chatbot was not intended to provide legal advice.
Speaking to WIRED, WhatsApp said its AI support chatbot was not intended to provide legal advice. That is a reasonable position, given that AI models can hallucinate and give users inaccurate legal information, potentially creating further problems for both the user and the company.
However, Redact argues that there is a difference between avoiding legal advice and withholding information about options available to users.
According to the researchers, some instructions went beyond discouraging legal advice by telling the chatbot not to recommend contacting data protection regulators or referencing relevant privacy laws.
The discovery raises a question about where companies should draw the line between limiting an AI assistant's responsibilities and ensuring users receive accurate information about their options
Broader implications for businesses using customer-facing AI
For businesses deploying AI in customer-facing roles, the incident highlights the need to pay attention not just to what their chatbots say, but also to what they are instructed not to say.
This matters most when AI systems handle complaints, account closures, billing disputes, privacy requests, or other situations where users need to understand their options before deciding what to do next.
Restricting a chatbot from giving professional advice is reasonable. Still, businesses should be careful not to let those restrictions prevent it from providing basic, relevant information or directing users to the appropriate authority or professional.
For security and AI governance teams, the incident underscores the importance of auditing chatbot instructions, testing responses against real customer scenarios, and checking whether internal restrictions conflict with published policies or users' legitimate options.
A chatbot does not have to invent a false answer to fail a user. Sometimes, leaving out an important option is enough.
For businesses deploying customer-facing AI, the lesson extends beyond preventing hallucinations. Companies also need to examine what their systems are instructed to omit, because those omissions can shape users' decisions and undermine trust just as much as an incorrect answer.
Other news: Microsoft has unveiled new security controls that monitor AI agent behavior and restrict access when risks emerge.