WhatsApp Help Pages Hide an LLM Prompt: Don't Recommend the Regulator

⚡ The Gist in 5 Seconds - Hidden instructions addressed to language models have been discovered in WhatsApp's help pages. - The text is aimed at AI assistants and crawlers, not at regular page visitors. - The key condition: the instruction only triggers in systems that read the entire page, including content invisible to humans. ### 🔍 What Was Found An analysis was published on the redact.dev blog and discussed on Hacker News, where it scored 590 points and 168 comments. The issue concerns WhatsApp's help pages — hidden instructions for LLMs were found in them, telling the model not to recommend a certain regulator to users. The specific agency is not named in the wording. The mechanism is a classic prompt injection in web text: the directives are embedded in the help page's markup so that they remain in the HTML and enter the context of any AI assistant or crawler that pulls the entire page. For humans, this block is invisible (hidden via styles/positioning), so a visitor to the help section sees only ordinary support text, while the model also receives a directive on what and whom "not to recommend." According to redact.dev, such insertions were found specifically in WhatsApp's help materials. ### 💡 Why It Matters This is a vivid example of how companies are starting to design content not only for humans and search engines, but also for LLMs — effectively embedding control instructions for AI intermediaries into the content. For owners of assistants and developers of RAG pipelines, it's a reminder: any external text must be treated as untrusted input, otherwise someone else's page will silently override the model's behavior. For users, it's a signal that an AI assistant's answer about a service may not be neutral, but shaped by a hidden directive on the source's side.