Hidden text on WhatsApp's help pages told AI systems what not to say, including not to recommend a regulator. That is the finding of a Redact investigation published October 19, 2026, written up by Redact's Dan Saltman in a post on redact.dev. According to the post, the text sat in the pages' source, invisible on screen, and has since been removed. WhatsApp says it was a bug.
The story is small in scale and large in implication: it is a concrete look at how a big platform tries to steer its support bot, and what happens when those steering notes leak onto public pages that other AI tools can read.
Quick reference: what is known
| Question | Answer |
|---|---|
| Who found it? | Redact (Dan Saltman), first reported to have found the notes October 8, 2026 |
| How many pages? | 80 of 545 linked WhatsApp help-center pages, 102 hidden blocks |
| What was notable? | About 20 blocks; two stand out (account deletion, reporting and blocking) |
| Did WhatsApp respond? | Per Redact, WhatsApp told WIRED it was a bug in chatbot instructions, now fixed |
| Are the notes gone? | Per Redact, all 102 blocks were gone by 02:08 UTC October 18 |
| Was bot behavior tested? | No, Redact did not test the chatbot's answers |
| Did Facebook and Instagram have the same? | Redact checked 300 pages and found no similar notes |
What did the hidden text say?
Redact points to two pages. On the account deletion help page, for people who cannot delete an account because they lost their phone or number, the hidden text told the AI to give only the waiting periods. It also said not to suggest WhatsApp could delete the account manually and not to suggest using a different phone. It included the line "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority," as quoted by Redact.
On the reporting and blocking page, the hidden text addressed users who mention legal action, lawsuits, lawyers, courts, or litigation. According to Redact, it told the AI to avoid legal topics and not to cite or interpret data protection laws such as GDPR, LGPD, or CCPA, even when the user raised them first.
Those are policy choices for a support bot. A company can reasonably say a chatbot should not give legal advice. The sharper question is the first note, because a lost-number user with no way to delete an account is exactly the person for whom a data protection regulator is a real avenue. The instruction, as quoted, does not say "do not give legal advice"; it names regulators.
How was it found?
Redact says it read all 545 pages linked from the WhatsApp help center and saved copies. Hidden AI text appeared on 80 pages in 102 blocks, marked with a name containing "HelpContentGalaxyLLMInclude" and not visible on screen. Most were routine housekeeping for the bot. The post lists the roughly 20 notable ones in a table.
Redact's timeline, as published:
- July 12, 2026: an Internet Archive copy of the deletion page has no note.
- September 4, 2026: an archive copy has the note.
- October 8, 2026: Redact first finds the notes.
- October 9, 2026: Redact checks how the pages appear to crawlers using the names of OpenAI, Anthropic, and Perplexity. The notes did not appear to those requests.
- October 15, 2026: both headline notes are still live.
- October 18, 2026: all 102 hidden blocks are gone from all 80 pages.
- October 19, 2026: Redact publishes; its post says WhatsApp gave WIRED a statement.
The October 9 crawler check is worth pausing on. If the notes were served to ordinary visitors but not to AI crawler user agents, outside AI tools would not have been reading them. Redact says it has not confirmed whether outside assistants followed the notes at all.
What did WhatsApp say?
Per Redact, WhatsApp told WIRED the notes were instructions for its support chatbot and appeared because of a bug, which it says is now fixed. A spokesperson said the company wants its chatbots "focused on providing the best advice to users about WhatsApp" and "not to offer legal advice." We could not retrieve WIRED's own article in this run, so that statement is attributed to Redact's account of it; check WIRED's original for the exact wording.
What is claimed versus verified
| Claim | Status |
|---|---|
| Hidden AI-directed text existed on WhatsApp help pages | Reported by Redact with archive copies as evidence; WhatsApp does not dispute it |
| The text told AI not to recommend regulators | Quoted by Redact from the page source; WhatsApp says it was chatbot instructions |
| It was a bug | WhatsApp's explanation, relayed via Redact; not independently verified |
| The chatbot actually withheld regulator information from users | Not tested; no evidence either way |
| Outside AI assistants obeyed the notes | Not confirmed |
| Other languages or pages outside the help center had notes | Not checked |
You can check the first claim yourself on archive copies of the two pages linked above, comparing versions from July and September 2026. Redact's post is the best source for the exact markup name to search for.
Why this is more than a footnote
It shows how support bots are really steered
Companies configure support assistants with long system prompts. This episode suggests at least one of those prompts was authored as page text and shipped inside the content the bot reads, with a marker to include it for the model and hide it from people. When a bug exposes that layer, the public sees the policy behind the polite answer.
It sits in the same family as hidden-text attacks
A hook catching a blank instruction card, showing how hidden text on web pages can steer AI systems
The mechanism is the same one attackers use: text invisible to humans but visible to machines. Our explainer on indirect prompt injection covers how hidden page text can steer an agent. Here the author is the site owner, not an attacker, but any agent browsing the page would still ingest the notes. A related study of hidden instructions in web traffic is summarized on AgentBeam, the agent security platform from the explainx.ai team, which is the place to look for how to detect and block hidden instructions before an agent acts on them.
It lands in a week of Meta data-and-AI scrutiny
Meta's AI-and-WhatsApp story has been busy: see our coverage of Meta AI on WhatsApp in Brazil's election campaign, Meta's October 30 crawling terms for publisher sites, and the earlier WhatsApp username privacy changes. For Meta's own personal-agent product, see Is Meta Muse safe?.
The regulatory angle
The GDPR gives people a right to complain to a supervisory authority. A bot told never to mention that route is a design choice a European data protection authority could take an interest in, but nothing in Redact's post says any regulator has opened an inquiry, and we found no such report. Treat the regulatory exposure as a question, not a finding. The distinction between "do not give legal advice" and "do not tell people they can complain to a regulator" is the one a regulator would likely examine.
What should you do?
If you use WhatsApp support:
- Do not rely on a chatbot alone for account deletion or data-rights problems. The official help pages describe the process, and your national data protection authority's website lists how to file a complaint.
- If a bot refuses to discuss legal options, that is not evidence you have none. Ask for a human agent.
- For data-rights requests under GDPR, LGPD, or CCPA, make them in writing and keep records.
If you build support bots or AI-readable sites:
- Do not put system-prompt-style instructions in public page markup. Anything on the page is readable by every crawler and agent that fetches it, whatever the CSS says.
- Separate the steering layer from the content layer. Keep bot policy in server-side configuration and ship only the user-visible article.
- Audit what your pages say to machines. Fetch your own pages with a plain HTTP client and an AI-crawler user agent, then diff against the rendered view.
- Write refusal rules that do not obstruct users' rights. A blanket "no legal topics" rule is easier to defend than a rule that blocks pointing to regulators.
- Log and review bot answers on rights-sensitive topics such as deletion, access, and complaints.
For a survey of tooling that watches agent behavior, see our AI agent security platforms overview.
What we do not know
The picture is incomplete in several ways. Redact did not test the chatbot, so we do not know whether it withheld regulator information in practice. We do not know whether other languages or pages outside the linked help center had notes. We do not know how long the notes were on every page, only the archive snapshots Redact cites. And WhatsApp's "bug" explanation has not been independently confirmed. We will update this post if WhatsApp or a regulator adds facts.
Details reflect Redact's published investigation and WhatsApp's reported statement as of October 20, 2026.