WhatsApp 帮助页面泄露 Meta 聊天机器人隐藏指令

Dozens of WhatsApp web pages designed to answer people’s questions about Meta’s encrypted messaging app inadvertently exposed the company’s own instructions to its chatbot. These included some sections titled “WHAT NOT TO SAY” to users asking questions and explicit instructions not to provide people with legal advice.

A new analysis has found that around 80 help and FAQ pages across WhatsApp’s website exposed text instructions that were designed for a large language model, specifically one of Meta’s own. The instructions, according to findings from privacy company Redact, appeared in the webpages’ code and contained written details that were not explicitly on the WhatsApp help pages.

Among the 100 blocks of hidden text, the directions appeared to provide guidance to steer an LLM from answering users questions in unhelpful ways and provide a glimpse at how instructions for chatbots are created. In one instance, the WhatsApp page on how to delete your account included instructions that, alongside other advice, said, “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority” about the issue. WIRED viewed the exposed text, which can still be seen when viewing the underlying code of archived versions of the help pages.

“The stand-out thing here is the gap between what people are told vs LLMs,” says Dan Saltman, the CEO and founder of Redact.dev. He says the company regularly scans the sites and privacy policies of big tech firms, and that many of the WhatsApp instructions the team found were straightforward and point people towards other official support channels. Still, it does not appear the instructions were meant for public exposure.

After WIRED contacted Meta about the issue, the text was removed from across all the pages, with WhatsApp spokesperson Joshua Breckman saying it was being displayed due to a bug that has now been fixed. “These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI,” Breckman says. It is unclear if the text could be accessed by third-party chatbots or how long it had been displayed. “We train our support chatbots to be as helpful as possible,” Breckman says.

The exposed instructions appeared under code blocks with a heading: “HelpContentGalaxyLLMInclude.” They are no longer visible in the live pages’ code.

On the page “About reporting and blocking on WhatsApp,” the hidden instructions included a series of statements to follow if a “user mentions legal action, lawsuits, lawyers, attorneys, courts, or litigation.”

“Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly,” other parts of the instructions said.

Breckman, the WhatsApp spokesperson, says that the company provides users with transparent information about their data rights and how they can take actions. “We want our product support chatbots to be focused on providing the best advice to users about WhatsApp—not to offer legal advice,” Breckman says.

Other exposed instructions said what to respond with if a user says they are deaf or hard of hearing, including: “This statement is background about the user, NOT their support request.” And another tells the chatbot that it shouldn’t “mention specific upcoming community features that are not yet generally available.”

Introducing the app: More ways to make the most of WIRED

Meta failed to catch hundreds of AI child abuse ads

Big Story: The search for Silicon Valley’s most powerful woman

Hackers got inside a Flock camera—its data shows how the system really works