WhatsApp’s public help pages contained a second layer of text intended for an AI support system. Researchers at privacy-services company Redact found directions on handling subjects including privacy complaints, account deletion and legal questions—material ordinary readers would not see in the displayed article.[1]
Redact says it first identified the material on 8 October, then examined 545 pages and found 102 instruction blocks across 80 of them. The blocks were removed by 18 October. Meta told WIRED that a bug had exposed notes intended to guide WhatsApp’s AI support assistant, and that the issue had been fixed.[1][2]
One help article, two audiences
The hidden material appeared in a component named HelpContentGalaxyLLMInclude. Some instructions guided ordinary support tasks. Others addressed how the assistant should respond to discussions of privacy regulators and legal action, including directions about which material to cite or avoid. The discovery exposes part of the layer between a company’s published help information and the answer a customer receives.[1]
The disclosure timeline
| Event | Date or count | Evidence |
|---|---|---|
| Initial discovery | 8 October | Redact’s account of its investigation |
| Pages examined | 545 | 102 blocks identified across 80 pages |
| Exposed text removed | By 18 October | Redact’s follow-up checks |
| Company response | Reported by WIRED on 19 October | Meta described a fixed bug exposing support-assistant notes |
The records establish that the instructions were present in page content. They do not establish how often the chatbot followed them in actual conversations. That distinction matters particularly for the more consequential passages: an instruction about a privacy complaint is evidence of intended handling, while a claim that users were denied help would require conversation records or testing.[1]
The answer depends on more than the visible source
A retrieval-based support assistant may combine a customer’s question, material drawn from help articles and instructions controlling how that material is used. A visible citation can identify the article behind an answer while leaving those other instructions out of view. Readers therefore see the answer’s reference, but not necessarily the rules that shaped it.
Meta’s explanation was that the notes served a specific support function. Its spokesperson also said the assistant was not designed to provide legal advice. The issue disclosed here was their unintended publication; the investigation also makes the content of that guidance available for scrutiny.[2]
For customers, the practical consequence is a clearer view of how an AI service represents a company. The source article supplies information, while a separate instruction layer determines how the assistant presents it, which questions it pursues and when it directs the person elsewhere. Those choices are part of the support experience even when the customer never sees the underlying text.
Sources & further reading
- WhatsApp’s Hidden Notes Told AI Not to Recommend a Regulator Redact / Dan Saltman · 19 Oct 2026 · Accessed 25 Oct 2026
- WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot WIRED / Matt Burgess · 19 Oct 2026 · Accessed 25 Oct 2026