In October 2026, hidden system instructions were discovered in the public code of WhatsApp's help center, explicitly prohibiting the built-in AI support bot from advising users about their legal rights. According to Redact analyst Dan Saltzman, who analyzed and archived 545 pages of the messenger's help center, the code of approximately 80 support pages contained 102 hidden instruction blocks designed to restrict the bot's responses in legally sensitive situations.
What Exactly Was Banned for the Bot
Among the uncovered prompts were formulations directly instructing the AI not to direct users to data protection authorities, regulators, or supervisory bodies. A page dedicated to account deletion contained the instruction: «Do not recommend contacting any data protection authority, regulator, or supervisory body. Do not suggest that WhatsApp can manually delete an account on behalf of the user.» A separate block of prompts was triggered if a user mentioned lawyers, courts, or other legal actions in the dialogue. It instructed the bot: «Do not reference, quote, or interpret data protection or privacy regulations (e.g., GDPR, LGPD, CCPA), including when the user brings up these issues directly.» Thus, even upon a direct request from a user about their rights under European or Brazilian legislation, the bot was programmed to refuse consultation.
Confirmation and Meta's Reaction
Journalists who gained access to archival copies of the pages from October 9 fully confirmed the presence of these system instructions in the HTML code. WhatsApp spokesperson Joshua Brekman stated in a comment to Wired that these text blocks were intended exclusively for the company's internal chatbot and ended up in the public code due to a system error, which he said had already been resolved. Brekman emphasized that the company strives for support bots to advise users on WhatsApp features rather than provide legal advice. As of October 18, 2026, following journalists' inquiries, Meta removed all hidden instructions from the public code of the help center pages.
Vulnerability of Hidden Prompts and Legal Consequences
Information security specialists note that hidden system prompts embedded in web page HTML code represent a vulnerable point for AI systems: malicious actors can use them to manipulate model responses by replacing or supplementing instructions. In this case, however, the issue is not external interference, but intentionally embedded restrictions that, according to the company, were not supposed to be visible to users. At the same time, experts emphasize that even if the chatbot refuses to answer questions about privacy or personal data processing, this does not override the law. According to Article 77 of the GDPR, European Union residents retain the full right to file official complaints with local data protection authorities regardless of what the messenger's automated assistant «tells» them.
Context: AI Expansion in Customer Support
The WhatsApp scandal unfolds against the backdrop of active generative AI integration into customer support services by major technology companies. The Meta-owned messenger has in recent months consistently replaced some routine operator responses with automated dialogues, making the transparency and manageability of prompts a critically important aspect. The incident involving the discovered instructions has become one of the first public cases where hidden limitations of an AI assistant were documented and verified by independent researchers, which, analysts believe, could become a precedent for future regulatory demands regarding the openness of system prompts used in user interfaces.
