WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot
An analysis by privacy firm Redact found that roughly 80 WhatsApp help and FAQ pages contained hidden text blocks—about 100 in total—intended for a large language model used by Meta. The embedded instructions, visible only in the page source, guided the chatbot to avoid recommending contact with data‑protection authorities, giving legal advice, citing regulations such as GDPR, or mentioning…
Key points
- Around 80 WhatsApp help/FAQ pages showed hidden LLM instructions, about 100 text blocks in total.
- Instructions told the chatbot not to give legal advice, cite regulators, or reference upcoming features.
- Meta said a bug caused the exposure, removed the text, and said the prompts are for its support chatbot only.
Meta said the hidden content appeared because of a bug and has removed it from all pages. WhatsApp spokesperson Joshua Breckman explained the text was meant for the company’s support chatbot to keep responses accurate and not provide legal counsel. It remains unclear whether third‑party bots could have accessed the instructions or how long they were exposed. Dan Saltman, CEO of Redact.dev, noted the gap between user‑facing information and the prompts steering the LLM.
The incident highlights how large tech firms embed covert prompts to shape AI behavior, raising privacy and compliance questions for regulators and users alike.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- US prosecutors allege ex-OpenAI employee had Saudi royal family ties in espionage case · 2 src
- Trump administration bans DeepSeek from civilian government systems · 2 src
- Anthropic pitches Congress frontier AI safety bill with $1B damage, 50-casualty thresholds · 20 src
- Human Technology Institute calls for stronger workplace privacy laws in Australia · 1 src
- AICPA and AIUC launch joint SOC 2 and AIUC-1 certification for AI systems · 1 src
Comments
via GitHub Discussions