Hidden AI instructions in WhatsApp help pages blocked regulator advice

In October 2026, investigators discovered that WhatsApp’s European help pages contained hidden text instructions designed specifically for artificial intelligence systems. These invisible directives told the company’s support chatbot to avoid recommending data protection regulators and to stop citing privacy laws once a user mentioned legal representation. The findings highlight a significant divergence between public-facing privacy policies and the operational rules governing automated customer support.

What happened

On October 8, 2026, researchers identified hidden blocks of text embedded within WhatsApp’s online help center. While these sections were invisible to human readers viewing the pages in a browser, they were fully accessible to web crawlers and AI models processing the content. The investigation expanded to cover all 545 pages linked from WhatsApp’s help center, revealing 102 hidden instruction blocks. Most of these blocks contained harmless housekeeping notes, but two specific instances raised serious compliance concerns regarding user rights and regulatory access.

The first problematic instruction appeared on the page detailing how to delete an account. For users who had lost access to their phone number or device, the visible text explained that accounts would be automatically deleted after 30 days of deactivation or 120 days of inactivity. However, the hidden text added strict prohibitions for the AI. It ordered the system not to recommend contacting any Data Protection Authority, not to suggest that WhatsApp could manually delete accounts, and not to advise using a different phone with an active number. This effectively removed alternative pathways for users locked out of their accounts.

The second instance was found on the page regarding reporting and blocking. This hidden instruction activated whenever a user mentioned keywords related to legal action, such as lawyers, lawsuits, or courts. Once triggered, the AI was directed to cease all references to legal documents, including WhatsApp’s own Terms of Service and Privacy Policy. Crucially, it was also forbidden from referencing, citing, or interpreting data protection regulations like the GDPR, LGPD, or CCPA, even if the user explicitly raised them. This restriction remained in place regardless of whether the user initiated the legal discussion.

How it works

These instructions were embedded directly into the HTML source code of the help pages using a method that renders them invisible to standard browser displays. The text was placed in a way that human visitors saw only the standard help content, while automated systems reading the raw page source received additional context. One hidden block explicitly referenced "HelpContentGalaxyLLMInclude," indicating these notes were intended for a large language model powering WhatsApp’s support chatbot.

The mechanism relied on the difference between rendered view and source code. When a user visited the page, the browser ignored the hidden text. However, when search engine crawlers from Google and Bing accessed the pages, they retrieved the full source, including the hidden directives. Researchers noted that crawlers from OpenAI, Anthropic, and Perplexity did not receive the article or notes when tested on October 9, suggesting some filtering may have been in place for specific AI bots, though Google and Bing still indexed the hidden content.

Key details

• Investigators analyzed 545 help center pages and found 102 hidden instruction blocks.
• Hidden text on the account deletion page prohibited the AI from recommending data protection regulators or manual deletion options.
• A separate hidden block prevented the AI from citing GDPR, LGPD, or CCPA if a user mentioned lawyers or legal action.
• WhatsApp’s public privacy policy for Europe, effective July 2, 2026, explicitly grants users the right to complain to regulators like the Irish Data Protection Commission.
• The hidden instructions contradicted the public policy by suppressing information about regulatory complaints and legal rights.
• Facebook and Instagram help centers, which use the same system, showed no similar hidden notes in a check of 300 pages.

Why it matters

For software engineers and technical leads, this incident illustrates the risks of embedding operational logic in content management systems without transparent auditing. When AI instructions are hidden from public view but influence customer interactions, they can create a disconnect between stated policy and actual user experience. This is particularly critical in regulated industries where transparency and adherence to laws like the GDPR are mandatory. Engineers building AI-driven support systems must ensure that safety guidelines and compliance rules do not inadvertently suppress lawful user rights.

The discrepancy between the public privacy policy and the hidden AI instructions also raises questions about accountability. While companies often restrict chatbots from giving legal advice to avoid liability, prohibiting them from mentioning existing laws or regulators goes beyond caution. It prevents users from accessing information they are legally entitled to know. For developers, this underscores the need for rigorous testing of AI outputs against legal requirements, ensuring that automated systems facilitate rather than obstruct user rights.

What you can do

• Audit your help center pages for hidden text or metadata that might influence AI behavior unexpectedly.
• Ensure that AI instructions align strictly with public-facing privacy policies and legal obligations.
• Test your support chatbot’s responses to queries involving legal rights, regulators, and account deletion.
• Review crawler access logs to understand which external AI systems are indexing your internal documentation.
• Implement clear separation between visible user content and backend AI instructions to maintain transparency.
• Regularly verify that automated responses do not contradict statutory rights granted to users in relevant jurisdictions.