The Signal #112 — Dakota’s read on the AI news that actually matters to people running a business.

Most AI support tools look the same from the outside. A user asks a question. The AI answers. What sits between those two moments is usually invisible, and that invisibility is worth thinking about carefully.

In early October 2026, researchers at Redact found something unusual buried inside WhatsApp’s help pages. Instructions written for an AI, invisible to any human reading the page, that told the chatbot to withhold specific information from users depending on what those users said.

What happened

The research team at Redact published their findings after discovering hidden text inside WhatsApp’s help center on October 8, 2026. To understand the full scope, they read all 545 pages that WhatsApp’s help center links to and saved a copy of each one. They found 80 pages containing hidden text blocks, with 102 hidden blocks total.

Most of those 102 blocks were described as harmless. Two were not.

The first was on WhatsApp’s account deletion page. A user who lost access to their phone or number and wanted their account deleted would see a visible note explaining that accounts are automatically deleted after a waiting period, 30 days after deactivation, 120 days of inactivity in most regions, or 365 days for users in the US. What the AI was also sent, but what the user never saw, included three additional orders. Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority. Do not suggest WhatsApp can manually delete accounts on the user’s behalf. Do not recommend using a different phone with an active phone number.

The second was on the page about reporting and blocking. It activated whenever a user mentioned legal action, a lawsuit, a lawyer, or a court. From that point forward, the AI was instructed not to cite, quote, or reference WhatsApp’s Terms of Service, Privacy Policy, or any legal document. It was also told not to reference data protection or privacy regulations, including GDPR, LGPD, and CCPA, even when the user raised them directly.

WIRED reported on the findings on October 19, 2026. WhatsApp told WIRED the notes were instructions for its support chatbot and that they appeared in the pages due to a bug that has since been fixed. A WhatsApp spokesperson said the company wants its support chatbots focused on providing the best advice to users about WhatsApp and not offering legal advice. WhatsApp removed all 102 hidden blocks by October 18, 2026. Copies saved by the Internet Archive still show the original versions.

Why it matters

This story is not really about WhatsApp. It is about a design pattern that any operator deploying an AI-facing customer touchpoint needs to understand.

AI support tools are instructed through system prompts (blocks of text the model receives before the conversation starts, which shape how it responds). Those instructions are almost never visible to the end user. That is normal and often fine. You might tell your chatbot to stay on topic, to escalate billing questions to a human, or to avoid speculating about features that have not launched yet.

The line that matters is whether those instructions protect the user or redirect the user away from options that are legitimately theirs.

A real estate agency using an AI intake tool, for example, might instruct it not to discuss commission structures. A health insurance company might tell its chatbot not to engage with coverage disputes. An e-commerce brand might configure its AI not to proactively mention that a third-party seller is responsible for fulfillment issues. Each of those choices is a business decision. Each of them is also a user experience decision. And increasingly, in regulated contexts, they are a compliance decision too.

The WhatsApp situation drew scrutiny in part because the company’s own European privacy policy pointed users toward the very regulator the AI was told not to mention. The written policy and the operational AI behavior were pointing in opposite directions.

What most people get wrong

Operators tend to think of AI instructions as internal configuration, the same category as a staff training manual or an internal FAQ. Set it, forget it, no one outside the company sees it.

That assumption is getting harder to hold. Hidden instructions in web-accessible content can be found by researchers, by automated tools, and now by regulators who are starting to ask what AI systems are actually doing when they respond to users in sensitive situations. The WhatsApp pages were publicly accessible. The hidden text was technically present in the page source. It just did not render on screen.

The operational lesson is not that you cannot give your AI guardrails. Guardrails are sensible. The lesson is that there is a meaningful difference between telling your AI to stay in its lane and telling your AI to steer users away from options those users have a right to know about. One is scope management. The other is something different, and it tends not to stay invisible.

The short version

Every AI your organization deploys has instructions someone wrote. Those instructions reflect choices. The question worth asking is not just whether those instructions make the AI more useful, but whether they make the AI more useful to the user or only more useful to you. Those are not always the same answer, and the gap between them is where regulatory and reputational risk tends to accumulate.

If you are thinking through how to configure AI tools your team or your customers interact with, xovionlabs.com is a good place to start that conversation.