WhatsApp's help pages hid LLM instructions: dont recommend a regulator
redact.dev
October 19, 2026
10 min read
59/100
Summary
WhatsApp embedded hidden instructions for its support chatbot in help-center pages that were not visible to ordinary visitors, according to copies examined on October 8, 2026. The instructions included telling the chatbot not to recommend data-protection regulators to users unable to delete an account and not to cite or interpret the GDPR, LGPD or CCPA after a user mentioned a lawyer, lawsuit or court. WhatsApp removed the notes from live pages by mid-October. The restrictions conflicted with WhatsApp’s European privacy policy, effective July 2, 2026, which says users may request erasure of their information and complain to the Irish Data Protection Commission or another competent authority. The hidden instructions did not bar discussion of erasure requests, but told the chatbot not to suggest that WhatsApp could manually delete an account for a user. WhatsApp told WIRED that the notes were instructions for its own support chatbot. The notes appeared in 102 hidden blocks across 545 linked help-center pages. Most covered housekeeping, product guidance or safety issues, but several directed the chatbot to withhold process details or discourage alternative account-recovery and deletion routes. The notes were sent to Google and Bing crawlers with the pages; tests using crawler names associated with OpenAI, Anthropic and Perplexity returned pages without them. It remains unknown whether WhatsApp’s chatbot or external AI assistants followed the instructions.
Key Takeaways
- WhatsApp removed hidden help-page instructions that told its support chatbot not to recommend data-protection authorities in some account-deletion cases.
- Hidden instructions barred the chatbot from referencing, citing or interpreting the GDPR, LGPD or CCPA after users mentioned lawyers, legal action, lawsuits or courts.
- WhatsApp’s European privacy policy states that users can request erasure and complain to the Irish Data Protection Commission or another competent supervisory authority.
- The hidden blocks appeared across 102 sections of 545 WhatsApp help-center pages, although most concerned routine guidance, terminology or user safety.
- WhatsApp told WIRED that the hidden notes were instructions for its own support chatbot, but it is not known whether the chatbot followed them.
What the discussion said
The thread focused less on a mysterious hack than on what it means for a company to embed operational instructions for an AI support bot inside public help pages. Several commenters thought the surrounding rules make Meta’s explanation plausible: a support bot should not invent legal interpretations, concede liability, or masquerade as counsel. They also saw a real engineering upside in treating official support pages as a structured, first-party source of bot instructions, particularly when multiple documentation teams need to steer answers consistently. The accessibility-oriented guidance was cited as evidence that not every hidden instruction was self-serving. But the regulator instruction poisoned that charitable reading. Critics argued that refusing to direct users to a data-protection authority is not merely avoiding legal advice; it steers people away from a remedy the visible policy itself says they possess. That contradiction made the setup look like covert manipulation, whether or not calling it prompt injection is technically precise. Others pushed back that companies reasonably design their own support agents to represent the company’s position, and that the actual exposure was the instructions becoming visible to users and external crawlers. The sharp concern was therefore not AI guidance itself, but AI guidance quietly overriding user-facing rights information.
Where opinion split
The central fight is whether the leaked text was ordinary support-bot guardrailing or a covert attempt to suppress users’ legal recourse. Defenders say preventing an LLM from hallucinating legal advice or corporate liability is prudent, and public structured instructions can be more auditable than a private backend. Critics say directing the model away from regulators while the page tells users they may complain to one turns a support bot into a rights-suppressing filter.
Coverage & discussions
Hacker News · redact.dev
Community discussions
Community Sentiment
Positives
- Keeping an AI support agent away from legal conclusions and liability claims can prevent confident hallucinations from creating false assurances or accidental corporate commitments.
- Publishing structured first-party instructions alongside support material could make bot behavior easier to coordinate and audit than burying every rule in an opaque internal API.
- Guidance that recognizes accessibility needs shows the instruction layer can help the bot interpret requests correctly instead of mechanically misclassifying them.
Concerns
- Telling the support model not to suggest regulators looks like steering users away from a formal remedy, especially when the public-facing policy acknowledges that remedy exists.
- The mismatch between human-readable rights language and model-only restrictions makes the chatbot seem designed to give a narrower answer than the documentation promises.
- Embedding hidden model directives in public pages risks contaminating third-party AI systems that crawl them, turning ordinary documentation into an unintended cross-model instruction channel.
- Calling the exposure a bug drew skepticism because the troubling issue for critics is the policy encoded in the bot, not merely the fact that users could inspect it.
