The help page had a second audience
WhatsApp’s help pages harbored a secret audience. Redact, a privacy and data-deletion firm, discovered machine-targeted text hidden in the page source under the identifier HelpContentGalaxyLLMInclude, invisible to ordinary visitors. This wasn't merely a stray note; it represented a strategic layer of communication.
Researchers audited all 545 linked help pages and found 102 such hidden blocks across 80 unique pages. A parallel audit of 300 Facebook and Instagram pages, operating on the same content management system, revealed no equivalent embedded instructions. The scale suggested deliberate, not accidental, implementation.
Why does placement matter? Support systems often retrieve help content as contextual information for AI models. Instructions embedded directly into this documentation can thus influence an AI's behavior and responses, even when those directives are completely hidden from the customer seeking assistance.
These hidden notes were not benign. While some were harmless, like clarifying “I am deaf” refers to a person, not an audio fault, others contained highly sensitive directives. This covert prompting mechanism allowed WhatsApp to steer AI responses without user knowledge, creating a shadow layer of policy.
The bot was told to steer around regulators
Instructions to the AI weren't always so benign. One directive, found on the "Delete your account" page, explicitly told the bot: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." It also forbade suggesting WhatsApp could manually delete an account.
Another alarming instruction appeared on the reporting page. If a user mentioned a lawyer, lawsuit, or court, the model was told not to cite or interpret GDPR, LGPD, or CCPA—even if the user introduced those laws directly. The implication was clear: steer users away from legal avenues and regulatory bodies.
These constraints stand in stark contrast to other, more innocuous guidance also found in the HelpContentGalaxyLLMInclude block. For instance, one note clarified that "I am deaf" describes the user, not an audio fault, while another directed the AI never to advise entering a false age.
Redact's audit revealed 102 such blocks across 80 pages, a mix of operational guidance and more consequential directives. WhatsApp later told WIRED these were instructions for its internal support chatbot, exposed due to a "technical bug." The notes were purged by October 18, 2026, but the Internet Archive preserves copies.
Public privacy promises, private AI guardrails
WhatsApp’s European privacy policy, a public document, explicitly names GDPR rights and the option to complain to the Irish Data Protection Commission. This stands in stark contrast to the hidden directive telling the support bot to avoid mentioning any Data Protection Authority or regulator. The dissonance raises eyebrows, suggesting a deliberate attempt to steer users away from formal recourse.
Meta spokesperson Joshua Breckman told WIRED the exposed instructions were intended for WhatsApp Support with AI, calling their public HTML exposure a "technical bug" that has since been patched. Breckman maintained the bot provides product help, not legal advice. This distinction, however, does little to assuage concerns about transparency.
While avoiding inaccurate legal advice is prudent, a blanket instruction to never mention complaint channels crosses a different line. This isn't just caution; it's suppression of information, raising serious questions about governance and user autonomy. For a deeper dive into the specifics, see WhatsApp's Hidden Notes Told AI Not to Recommend a Regulator.
Redact's findings, discovered on October 8, 2026, revealed 102 hidden blocks across 80 distinct pages out of 545 audited help center pages. These instructions, tagged under HelpContentGalaxyLLMInclude, were served to search engine crawlers like Google and Bing, but not specific AI user-agents. By October 18, 2026, WhatsApp had purged all 102 blocks, though the Internet Archive retains copies.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The notes vanished; the unanswered questions didn’t
Redact reported its discovery on October 8, 2026. The most problematic instructions, verified live on October 15, vanished by October 18, 2026, when WhatsApp purged all 102 hidden blocks from its help pages. Archived copies persist, allowing independent verification.
This swift deletion raises more questions than it answers. Researchers did not prove that WhatsApp’s production chatbot ever followed these specific directives, nor could they ascertain if similar rules remained in server-side prompts. The exposed instructions were merely one layer of control in a complex AI system.
This incident offers a crucial lesson for AI teams and users alike. Retrieved documents can carry hidden directions, silently shaping bot behavior. Companies must rigorously audit what their AI models ingest, disclose meaningful constraints, and ensure support answers align with published user rights. The stakes involve trust, compliance, and the integrity of customer support in the AI era.
Frequently Asked Questions
What did researchers find in WhatsApp’s help pages?
Redact found 102 hidden instruction blocks across 80 of 545 WhatsApp help pages. The notes were present in the page source, not visible on screen.
What did the instructions say about regulators?
One note told the AI not to recommend contacting a Data Protection Authority or other regulator. Another said not to cite privacy laws in certain legal discussions.
Did WhatsApp’s chatbot follow the hidden instructions?
Researchers did not establish whether the production chatbot followed the notes in customer conversations. WhatsApp said they were intended for its AI support service and that their exposure was a technical bug.
Are the hidden notes still on WhatsApp’s help pages?
The reported notes were removed from the live pages by October 18, 2026. Archived copies remain available through the Internet Archive.