Can't even keep its own secrets! Grok exposed in data leak—why do AI giants always stumble on privacy?

Beating AI News Flash: Security firm Claroc has discovered that 35,578 apps on Grok Build can be accessed without logging in, some of which involve corporate secrets and personal privacy.

Even Musk's own companies were affected. Twenty-four apps created by SpaceXSI and X employees exposed internal sales plans, customer data, and advertising proposals. Tesla employees' performance and bonus information was also made public, along with even ICU patients' examination reports and employees' plaintext passwords.

Claroc estimates that approximately 650 to 800 apps contain private data. Most of the rest are just ordinary mini-games, calculators, and the like.

The problem lies in how apps are shared. Users have Grok build a tool, click publish, and a webpage is generated. But Claroc found that the publishing interface defaults to "anyone with the link can access."

Although Grok adds random words to URLs to make them hard for others to guess, these URLs get recorded by third-party web services. Researchers simply queried these records and found tens of thousands of apps in bulk.

It's like leaving the door unlocked and just hiding the house number. This kind of problem has repeatedly occurred at multiple AI companies.

In 2025, some of ChatGPT's shared conversations were indexed by Google, and OpenAI subsequently withdrew the feature allowing users to proactively enable search indexing. In July of this year, some of Claude's publicly shared conversations and apps also appeared in search results. In October, Claroc found over 6,800 shared chat URLs from Meta Muse, some containing corporate secrets, and Meta subsequently suspended the related sharing feature.

The specific causes of these incidents are not entirely the same, but they all expose one problem: users think they are only sending links to a few people, but in reality, this content may be found by strangers across the entire internet.

Grok has another layer of risk. Claroc asked it to build a customer system for a four-person sales team. Although the AI added Google login, it did not properly check who was eligible to join the team. Researchers used an account that had never been invited and successfully viewed, modified, and deleted customer data.