What did researchers find in apps built with Grok Build?
That nearly all the live apps they examined were open to anyone. Vibe coding means building software by describing it in plain language to an AI, without writing code yourself. Grok Build, a tool from SpaceXSI (formerly xAI), does exactly that, then publishes the app straight from the chat. Security firm Claroc identified 43,601 app names on the grok.me domain. Of the 37,571 live apps it examined, 35,578, or 94.7%, could be opened without authentication, meaning nobody was asked who they were.
According to Claroc, most of the open apps were simple games, calculators or landing pages. But the researchers estimate that between 650 and 800 apps held private data. They included a hospital directory of more than 3,500 names with patient information, a transport company's recruitment database covering more than 1,000 people, and passwords stored in plain text, some of them linked to admin accounts. The researchers also say they found 24 open apps built by staff at SpaceXSI and X.
SpaceXSI says apps stay private until their creator shares them. When publishing, the creator chooses access for anyone with the link, for invited users, or for themselves only. The researchers point to those sharing settings: a choice made in a few seconds decides who can see the data.
What are the main vibe coding security risks?
The AI builds what you ask for. The risks live in what doesn't show up on screen:
- User login. Who can open the app? A hard-to-guess web address protects nothing: Grok Build addresses usually consist of four random words, and a script can try combinations until it hits real ones.
- Access rights. Once logged in, who sees what? In their tests, Claroc's researchers edited a shared contact database without logging in, and a Google account from outside the team was able to open another app.
- Personal data. Where are names, addresses and passwords stored, and are they encrypted, meaning unreadable without a key?
- Backups. If data is deleted by mistake, or a change breaks the app, can you roll back?
The code itself is not always safe either. Security vendor Veracode gave 80 coding tasks to more than 100 AI models: in 45% of cases, the code they produced contained a security flaw (study published in July 2025). Without reading the code, you can't tell whether your app is one of them.
Do new hosting platforms for AI-built apps fix the problem?
Only partly. Automattic, the company behind WordPress.com, opened Spacefast on 9 November 2026. It hosts sites and apps built with ChatGPT, Claude, Cursor or Grok, and gives each project hosting, a custom domain and a database.
Its defaults are cautious. A project is visible only to its author until it is shared. The AI agent has to ask for your approval before sensitive actions, such as rolling back to an earlier version or changing DNS settings (the settings that point your domain name to your site). Every update creates a version you can return to.
However, built-in user login for the apps themselves has only been announced as coming soon, with no date, according to Blog du Modérateur. And no platform can decide for you which customer should see which records.
Who is responsible if your customers' data leaks?
The business that collects the data, even if an AI wrote the app. Under GDPR, that business is the data controller: your company, as the client of the tool or the service provider. When personal data is lost, altered or accessed without authorisation and this puts people at risk, the controller reports the breach to the data protection authority (the CNIL in France) without undue delay and, where feasible, within 72 hours of becoming aware of it. A late notification must explain the delay, and the details can be provided in phases.
The processor that hosts or maintains the app has a different duty: it must notify the controller without undue delay after becoming aware of the breach (GDPR, Article 33(2)). Reporting to the authority remains your responsibility.
Every breach must also be logged internally: what happened, which people and data were affected, the likely consequences and the steps taken. When the risk is high, the people affected must be told too. For a young SaaS, that means emailing your first customers to say their data was exposed.
When is an AI-built prototype enough, and when do you need a real first version?
A prototype is enough as long as nobody puts real data in it. Use it to test an idea with potential customers: show them screens and note their reactions. Fill it with made-up data and only let in people you invite.
Once customers sign up, pay or enter personal information, you need a real first version. The researchers' advice, as reported by Le Monde Informatique, is to enforce authentication, manage permissions, check the data being handled and review security before going live. Put these questions to whoever builds your app, yourself included:
- does every page that shows data require a login?
- does a logged-in customer see only their own data, never someone else's?
- are passwords stored in something other than plain text?
- do you know where the data is hosted and how to restore it?
- has someone who can read code reviewed the sensitive parts before launch?
To launch a SaaS with us, you start with a first module limited to 1 or 2 features, at a fixed €6,000, live in 4 weeks. It covers front-end and back-end development (the screens and the server side), testing and deployment. Week three is for fine-tuning: what happens when customers make mistakes, skip a field or click three times. If you already have an AI-built prototype, we can take it over after a code audit: we tell you what can stay, what needs rewriting, and why.