WhatsApp’s Hidden AI Instructions Exposed in Help Pages, Then Removed

Meta quietly exposed internal chatbot instructions inside dozens of WhatsApp help pages, including prompts that told the company’s AI support system not to give legal advice and not to suggest users contact regulators. The text was removed after WIRED flagged the issue, highlighting how easily guidance meant for large language models can leak into public-facing webpages.

The disclosure matters because it offers a rare look at the hidden rules shaping AI-powered customer support at one of the world’s largest messaging platforms. It also raises fresh questions about transparency, legal boundaries, and the risks of publishing chatbot instructions in web code that can be viewed by anyone who knows where to look.

What was exposed on WhatsApp’s help pages?

About 80 WhatsApp support and FAQ pages contained hidden instructions intended for Meta’s chatbot rather than for human readers. Those instructions were embedded in the pages’ underlying code and were not meant to appear on the site itself.

The material was identified by privacy-focused company Redact, which said it found roughly 100 blocks of hidden text across those pages. The instructions were attached to support content on WhatsApp’s website and appeared to steer an AI assistant on how to respond to certain kinds of questions.

In archived versions of the pages, the hidden text could still be seen in the code even though it was never shown prominently to users visiting the live site. After WIRED contacted Meta, the company removed the text from the pages.

How did the hidden instructions work?

The instructions were placed inside code blocks labeled “HelpContentGalaxyLLMInclude,” a naming convention that strongly suggests they were designed for a large language model. In practical terms, they appear to have functioned as policy guardrails for Meta’s support chatbot.

Many of the prompts were straightforward and operational, telling the AI where to send people for help and what not to discuss. Others were more specific, offering guidance about sensitive topics such as legal disputes, privacy laws, disability-related support, and unreleased product features.

According to Redact founder Dan Saltman, the most notable issue is the disconnect between what users see and what an LLM is told behind the scenes. He said the hidden guidance offers a glimpse into how chatbot instructions are written and how different that internal logic can be from the public-facing support language.

“The stand-out thing here is the gap between what people are told vs LLMs,” Redact CEO Dan Saltman said, describing the exposed instructions as a window into how large companies shape AI support behavior.

Why did the legal instructions stand out?

The most sensitive material involved guidance for situations where users raise legal concerns. On a page about reporting and blocking on WhatsApp, the hidden text told the chatbot how to respond if someone mentioned lawsuits, lawyers, courts, litigation, or similar issues.

One of the clearest examples instructed the system not to cite or interpret WhatsApp’s Terms of Service, Privacy Policy, or any other legal document. Another directed the chatbot not to offer legal opinions, defenses, or interpretations on behalf of the company.

The instructions also said not to reference privacy or data-protection laws, including GDPR, LGPD, and CCPA, even when users brought those laws up themselves. In one case, the hidden text told the AI not to recommend contacting a data protection authority, regulator, or other supervisory body.

That is significant because legal guidance is exactly where support chatbots can create the most risk. If an AI system appears to discourage users from seeking outside remedies or from understanding their rights, it can raise concerns about fairness, consumer protection, and compliance.

What did Meta say?

Meta said the text was not meant for users and blamed a bug for making it visible in the pages’ code. A WhatsApp spokesperson said the company fixed the issue after being alerted and emphasized that the instructions were meant to improve the chatbot’s accuracy and relevance for users who message WhatsApp Support with AI.

The company also said it aims to keep support chatbots helpful while focusing them on WhatsApp-specific support rather than legal advice. Meta maintained that users receive transparent information about their data rights and how to take action if needed.

WhatsApp spokesperson Joshua Breckman said the text represented instructions for the chatbot rather than advice shown directly to users, and said the company fixed the bug after it was reported.

Meta did not say how long the instructions had been exposed, and it remains unclear whether third-party systems or automated tools could access them before they were removed.

What the leaked prompts reveal about AI support systems

The exposed text offers a rare look at a growing but often invisible layer of modern customer support: the rules that shape how chatbots talk, what they avoid, and when they should redirect people elsewhere.

As more companies use AI assistants on support pages, they are increasingly forced to codify boundaries. Those boundaries can cover everything from legal disclaimers to accessibility issues, safety concerns, product launch secrecy, and the handling of disputes.

The WhatsApp leak suggests that support chatbots are not simply trained to answer questions. They are also tightly constrained to avoid creating legal exposure, contradicting company policy, or encouraging users to pursue external channels that might be inconvenient for the business.

That does not necessarily mean the instructions were improper. In many cases, companies are entitled to limit the scope of automated support. But it does show the tension between automated assistance and user autonomy, especially when the questions involve privacy rights or formal complaints.

Why companies rely on hidden AI instructions

Companies hide these instructions for a simple reason: they are operational controls, not customer-facing content. LLM prompts often contain internal routing rules, moderation language, style guidelines, and safety constraints that are meant to stay behind the curtain.

In a support context, hidden instructions may tell a chatbot:

• which topics it can answer directly,
• when to refer people to a human agent,
• what legal claims it must avoid,
• how to respond to accessibility-related issues, and
• what product details are too early to share.

The problem is that web pages are often a messy mix of visible content and machine-readable code. If internal instructions are inserted poorly, they can become discoverable in source code, cached pages, or archived copies.

How the WhatsApp case unfolded

The sequence was straightforward: Redact found the hidden text, WIRED reviewed it, Meta was contacted, and the company removed the instructions from the live pages. The incident shows how routine website maintenance can become a public accountability issue when AI systems are involved.

The incident is unlikely to be the last of its kind. As more public websites rely on AI-generated or AI-governed support content, internal prompts can become accidental disclosures unless companies carefully separate visible content from machine instructions.

What specific examples were found?

Some of the hidden text was routine, telling the chatbot how to handle common support situations. Other examples were more revealing because they suggested the company wanted the bot to avoid certain answers, even when users were asking about rights or formal remedies.

One instruction said that if a user mentioned being deaf or hard of hearing, the chatbot should treat that detail as background information rather than the core support request. Another told the AI not to mention upcoming community features that were not yet broadly available.

Those examples show the wide range of concerns companies encode into AI support systems. Accessibility, release timing, legal exposure, and user escalation all sit side by side in the prompt layer.

How does this fit into broader AI accountability concerns?

It fits into a larger pattern: as AI systems become part of consumer-facing products, the rules governing them are often hidden from the people they affect. That creates a transparency gap between the public version of a service and the machine logic behind it.

Researchers and privacy advocates have argued that this gap matters because AI instructions can influence whether users receive help, whether they are steered away from complaints, and whether a company’s support system is effectively self-protective.

In this case, the most notable issue is not that WhatsApp had rules for its chatbot. It is that those rules were exposed in a way that made them readable by outsiders, revealing the company’s internal approach to risk management and user support.

What users should take away from the leak

For most WhatsApp users, the immediate practical impact is limited: the exposed text has been removed, and there is no indication that account security or message encryption was compromised by the incident alone.

Still, the episode is a reminder to be cautious when relying on AI support tools for issues involving privacy, legal rights, or formal complaints. Automated systems can be helpful, but they may also be designed to keep the conversation narrow, avoid liability, and push users toward company-controlled channels.

Anyone dealing with account deletion, data rights, or a serious dispute should consider preserving records, reviewing official policies directly, and seeking outside help where appropriate. A chatbot can explain procedures, but it should not be treated as the final authority on legal or regulatory questions.

Why this matters beyond WhatsApp

This is not just a WhatsApp story. It is a snapshot of how major technology companies are building AI into support workflows while trying to keep those systems aligned with legal strategy, product policy, and customer service goals.

The more AI is used to handle consumer interactions, the more important it becomes to understand who programs the bot, what instructions it follows, and how those instructions are tested, audited, and hidden from public view.

Leaked prompts may seem like a niche technical issue, but they can reveal a company’s priorities more clearly than polished marketing language. In this case, the hidden text showed a support system designed to be helpful, but also carefully constrained on anything that might invite legal scrutiny.

That tension is likely to define the next phase of AI customer support. Companies want speed and scale. Users want clarity and redress. Regulators are increasingly asking for transparency. Incidents like this show how hard it can be to balance all three when the instructions themselves are part of the problem.

Timeline of the WhatsApp prompt exposure

The episode underscores a simple reality of the AI era: the hidden rules behind automated systems can be just as important as the answers those systems provide. When those rules leak, they can expose a company’s priorities, limits, and legal caution in unusually direct fashion.