WhatsApp’s hidden chatbot notes show why fetched text needs boundaries

Redact reports finding 102 hidden instruction blocks across 80 WhatsApp help pages, including directions for a support chatbot to avoid recommending a regulator in an account-deletion scenario. The notes were later removed. According to the report, WhatsApp said a bug exposed internal chatbot instructions. Redact did not test whether any assistant followed them, so the finding concerns exposed instructions rather than demonstrated manipulation of model answers.

Redact · Oct 19, 2026

Why it matters

Content fetched for an assistant can contain instructions intended for another system, even when the visible page looks like ordinary documentation.

More context

The researchers describe comparing archived pages and inspecting hidden content. They also distinguish instructions sent to search crawlers from requests that returned no article to several AI crawler identities. An editorial next step is to test whether retrieval pipelines treat source instructions as data rather than governing policy.

Keep in mind

The current pages no longer contain the reported notes. The account is the researchers’ original investigation; it does not establish how long every note existed or prove downstream assistant behavior.

Original read