Building applications with a prompt, in minutes instead of weeks: that is exactly the promise behind the app features of modern AI assistants. That speed comes at a price is shown by a recent discovery by the security firm Claroc. It came across more than 35,000 apps created with Grok that could be accessed by anyone on the web without logging in.
Among the openly accessible applications, according to the report, there were also patient data as well as confidential internal documents of the provider itself. So this is not a marginal case involving hobby projects, but content that must under no circumstances be public.
The cause was not an attack, but a default setting
What is remarkable is how mundane the trigger was. When publishing an app, the access option Anyone with the link – i.e. access for everyone who knows the link – was preselected by default. Anyone who did not actively change the setting effectively placed their application on the open web.
This is a classic case of insecure defaults. Default settings are not a neutral technical side issue, but a decision by the manufacturer about what happens in case of doubt. And in case of doubt, what happens for most users is exactly what is preset. Anyone who can put together an app with a prompt in a few minutes, experience shows, does not read the sharing menu particularly attentively.
There is also a false security assumption: a long link that appears random feels secret. In reality, "not linked" is not a protection mechanism. Such URLs turn up in search engines, referrer headers, chat histories, ticket systems and automated scans – and it was precisely such scans that apparently found the apps.
Cleaning up is harder than publishing
How the discovery was handled illustrates a second problem. According to the report, the provider took the 24 apps of its own employees offline; the remaining applications stayed accessible. That is understandable because the content belongs to other accounts – and yet it is a dilemma: formally, the people and organisations who uploaded them are responsible for the data. Many of them presumably do not know to this day that their app is public.
For companies this means: nobody cleans up after you. If employees in a business unit independently create AI apps and store customer data, quotes or HR documents there, shadow IT with a direct external impact arises – without the IT department hearing about it.
Why prototypes are particularly at risk
Prototypes are created quickly, informally and often on the margins of official processes. That is exactly what makes them useful. It becomes risky at the point where a prototype quietly slips into everyday operations: someone shares the link with a colleague, the colleague with a customer, and suddenly real personal data flows through an application that has never seen a review.
With apps generated by prompt, this is exacerbated. The code is often not read, the data storage is not questioned, and there is no point at which someone consciously decides: this is going into production now. The transition from experiment to production system happens unnoticed.
Four points before going into production
For the step from AI prototype to deployed application, a manageable set of checks is enough – what matters is that they take place at all:
- Access concept: Who may see the application, who may change data? Being publicly reachable is a conscious decision, not a default state. Sharing based purely on knowledge of the link must be ruled out for internal data.
- Data protection review: Which personal data is processed, where is it stored, for how long? Health data and comparably sensitive categories have no place in a prototype in principle – test data is sufficient.
- Code and architecture review: Generated code is code like any other and needs a four-eyes principle: authentication, interfaces, handling of credentials, logging.
- Inventory and expiry date: Every AI app belongs on a list with a responsible person. Prototypes without an owner are switched off after a defined deadline instead of being forgotten.
Governance instead of prohibition
The obvious reaction – banning AI tools in the company – does not solve the problem, but shifts it to private accounts and thus completely out of sight. What makes more sense is a clearly described path: approved tools, rules on which data classes may be processed in them, and a simple way to move a prototype into the regular development and operations model.
This also includes regular monitoring of sharing settings. Platform default settings change with product updates; what was internal yesterday may be configured differently after a release. A recurring review of your own published applications costs little and prevents exactly the state that Claroc documented.
Assessment
For web and software projects, the task shifts: in future, the bottleneck will not be creating an application, but the decision whether and under what conditions it may be operated. Review, permissions concept and data protection review are therefore not a bureaucratic appendix, but the actual added value of professional implementation.
Anyone who permits AI prototypes should at the same time define where they end – and who switches them off when no one is responsible any more.