An AI agent is the right tool for drafting a business app fast, but access control, secrets and data storage belong in plain, deterministic code that a person reviews. The Grok Build leak of about 35,000 apps shows what happens when an AI coding platform's defaults decide who can see private data.
According to Ars Technica, security firm Claroc found that Grok Build share links, which were meant to be unguessable, could be enumerated. Opening a link usually exposed the app and all its data. The exposed data included medical exams of 3,784 workers, employee rosters, sales pipelines and confidential proposals. It also included 24 apps from SpaceXSI itself, the company that makes Grok. Those links were taken down only after Ars asked questions.
What this means for a small business
A vibe-coded dashboard for invoices or leads takes minutes to build. The risky part is the link you send a colleague. On many platforms "shared" quietly means "public to anyone who finds it," and Claroc says most other AI platforms share this flaw. In my builds, the LLM handles the fuzzy work, like reading documents or sorting requests. Logins, permissions and where the data lives are ordinary code on a server the client controls. Nobody prompts that part into existence.
Why it matters: Before real customer or employee data goes into an AI-built app, check whether its share link sits behind a sign-in. If it does not, keep test data only.
Source: Ars Technica