A post on redact.dev, picked up on Hacker News, claims that WhatsApp's help pages contain hidden text aimed at language models. The content, invisible to a human visitor but readable by an automated system parsing the page's source code, reportedly instructs AI tools not to recommend that users contact a regulator when they have a complaint.
This kind of technique resembles indirect prompt injection: ordinary web content is altered to include instructions meant to influence how an AI system behaves when it reads the page, whether that system is a browser-integrated assistant, a customer-support chatbot, or an agent summarizing the content on a user's behalf. The apparent goal here would be to stop such systems from suggesting that dissatisfied users escalate their complaint to a data-protection authority or similar regulatory body.
While the precise technical details of the finding remain limited based on what's publicly available, the case highlights a growing concern in the security community: as AI assistants are increasingly used to browse, summarize, or respond on users' behalf, web pages themselves become a new attack surface. Both companies and malicious actors could attempt to steer these systems' outputs by embedding hidden directives in otherwise ordinary content.
The episode also raises a governance question for large platforms: if such instructions genuinely exist, they blur the line between optimizing customer support and attempting to sidestep users' legitimate avenues for recourse. Meta has not publicly commented on the allegations as far as currently available information shows.