Eighty of 545 WhatsApp help pages carried text for a chatbot that people reading those pages could not see.
The privacy company Redact found 102 hidden instruction blocks across those 80 pages in an October audit of WhatsApp’s help centre. Most were ordinary support guidance. Two did something narrower and more consequential: they told the chatbot to withhold routes that WhatsApp’s own public policy names.
The full count keeps the finding in proportion. Among the other blocks, Redact found instructions that protected users, such as refusing workarounds for age checks and preserving chat history that might be evidence after a scam. Others clarified product terms or routed subscription questions to an app store. The presence of hidden chatbot guidance was therefore not the problem by itself. The test is what each instruction asked the model to withhold.
The instruction behind the deletion page
On the help page for people who could no longer access the phone or number tied to their account, Redact recorded this instruction: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.”
That is not a paraphrase of a cautious legal policy. It is a direction to omit a category of institution. The same block also told the chatbot not to suggest that WhatsApp could manually delete an account, leaving automatic deletion after 30 or 120 days, or 365 days for US users, as the route it should describe.
The wording has since disappeared from the live page. WIRED independently said it viewed the exposed text in archived page code and reported WhatsApp’s explanation: the material was intended for its support chatbot and appeared publicly because of a bug that had been fixed.
The bug explains why outsiders could read the instruction. It does not explain the instruction’s substance.
The second block went further
A separate block applied after a user mentioned legal action, a lawyer or a court. It began defensibly: do not give legal opinions, discuss liability or interpret legal documents on WhatsApp’s behalf.
Then it added a broader line: “Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly.”
The last clause matters. A chatbot can avoid interpreting a law while still naming the law or pointing to a regulator. This instruction grouped those actions together, even when the user introduced the subject.
WhatsApp told WIRED that it wants support chatbots focused on product advice rather than legal advice. That boundary is reasonable. The hidden text crossed it by suppressing references, not only interpretations.
The public policy points the other way
WhatsApp’s current privacy policy for the European region says users “always have the right to request access to, rectification of, and erasure of your information.” It also says users have “the right to lodge a complaint” with the Irish Data Protection Commission or another competent supervisory authority.
Those routes do not depend on a chatbot volunteering legal analysis. Article 17 of the GDPR establishes a qualified right to erasure, while Article 77 establishes the right to complain to a supervisory authority. Whether either route succeeds in a particular case is a legal question. Naming that the routes exist is a factual one.
The two texts therefore did not merely address different audiences. The public policy named a complaint route; the hidden support instruction said not to recommend it. The policy named data rights; the support instruction said not to reference the regulations behind them once legal action entered the conversation.
What the audit does not prove
The finding is about instructions, not observed chatbot behaviour. Redact says it did not test whether WhatsApp’s support chatbot followed the blocks, and it did not establish whether outside assistants obeyed text served with the pages. The audit also cannot date the full exposure precisely: an archived July page lacked the deletion instruction, while a September copy contained it.
The removal is easier to establish. Redact found the text on October 8; WIRED says Meta removed it after the outlet asked about the issue; the live deletion page no longer contains the regulator line. That confirms a before-and-after change in the page source, not the behaviour of the model consuming it.
Those limits rule out a claim that users were denied a regulator referral in practice. They do not erase the documented design choice. A company can prohibit its support model from giving legal advice without prohibiting it from naming the complaint channel printed in the company’s own policy.
The hidden blocks are gone. The unresolved question is whether the revised chatbot now keeps that distinction.