The Share Link Was Never the Lock: 35,578 Apps Wide Open
Claroc researchers tested 43,551 Vector Build app addresses in a browser with no account signed in. They found that 35,578 opened. That was 94.7% of the apps still online that day, according to the article. Only 1,993 addresses returned a 401 response, an error indicating that access was denied. The apps each had a hostname under Vector.me. The article explains that Claroc did not need to guess the random-word names. DNS lookups made when people first opened those hostnames were collected by passive DNS services, creating searchable datasets. The deeper failure was that reading an app did not trigger an authorization check. SpaceXSI said publishing involved several clicks and employee accounts defaulted to Only you. But Claroc used a paid account and found Anyone with the link selected; the researchers argued that this did not change the read-access failure. The article recommends checking permissions on each object and watching for bursts of anonymous denials, which can signal enumeration.
What did Claroc find when it opened Vector Build app addresses without signing in?
Claroc researchers tested 43,551 Vector Build app addresses in a browser with no account signed in. They found that 35,578 opened. That was 94.7% of the apps still online that day, according to the article. Only 1,993 addresses returned a 401 response, an error indicating that access was denied.
The apps each had a hostname under Vector.me. The article explains that Claroc did not need to guess the random-word names. DNS lookups made when people first opened those hostnames were collected by passive DNS services, creating searchable datasets. The deeper failure was that reading an app did not trigger an authorization check.
SpaceXSI said publishing involved several clicks and employee accounts defaulted to Only you. But Claroc used a paid account and found Anyone with the link selected; the researchers argued that this did not change the read-access failure. The article recommends checking permissions on each object and watching for bursts of anonymous denials, which can signal enumeration.
How many tested apps opened, and what share of the apps still online did that represent?
The central scale finding is that Claroc researchers could open tens of thousands of Vector Build app addresses without signing in. They tested 43,551 addresses, and 35,578 opened. The article reports that those open apps represented 94.7% of the apps still online on the test day.
The remaining result was not all the same: 1,993 addresses returned a 401, which means access was denied. Vector Build published each app at its own hostname under Vector.me. Claroc found hostnames through passive DNS datasets, rather than by guessing the four random words Vector proposed for each name. The permission failure occurred when the app data was read.
These figures describe the tested addresses and apps still online that day, not every app ever published. The article’s point is that difficult-to-guess hostnames did not provide access control. It urges platforms to check whether each requester may read each object, and to log and alert on anonymous denials that may appear in bursts during enumeration.
What kinds of private information could strangers see in some of the apps?
The article describes the exposed object as an app and the data behind it. It names databases, records, and files the app generated as examples. The risk matters because a person who could open an app address might also receive data that its owner expected to keep private.
The underlying issue was not simply that a link existed. Claroc found that authorization on read was missing, so possession of the reference could become the whole access-control system. The article’s example of a correct design fetches an artifact, checks whether the subject may perform the read action, and returns an error if permission is absent.
The article does not list particular personal records, credentials, or other specific contents found in the tested apps. It also does not say that every opened app contained private data. Its advice is to inventory reachable objects, test private links in a private browser window, and review logs for unauthenticated reads of objects that should not be public.
How could people find random-word app hostnames without guessing them?
Vector Build proposed four random words for each app hostname under Vector.me. The article says those words were genuinely unguessable, but that did not keep the hostnames secret. The key distinction is between guessing a name and discovering a name that has already been looked up.
When someone first opens a hostname, it is looked up in DNS. Passive DNS services collect those lookups and make them searchable. Claroc used a pattern search across those datasets to find hostnames, then tested the app addresses in a browser without signing in. The discovery step therefore did not depend on defeating the random-word naming scheme.
The article’s lesson is that a long or random reference can raise the cost of guessing but cannot replace a permission check. A hostname may become discoverable through an index even if it is hard to predict. Platforms should check authorization when an object is read and monitor anonymous denials, since enumeration can produce a burst of failed requests.
What is an insecure direct object reference, and how did missing read checks create one here?
An insecure direct object reference, or IDOR, is a flaw where a system exposes an object through a reference but fails to verify that the requester is allowed to access it. The object might be an app or its data; the reference might be a URL or identifier. The flaw is about missing authorization, not merely about whether a reference is easy to guess.
In the Vector Build case, apps had their own hostnames, and Claroc could find those names through passive DNS. The article says authorization on read was missing. As a result, possession of the link became the entire access-control system: a stranger could request the app and receive its data without a permission decision.
The article’s illustrative fix checks who is asking, which object they want, and whether they may read it. A denial should be logged, counted, and alerted on, especially for anonymous requests. The article calls the missing policy check the core architecture issue and recommends applying such checks at every object-reading surface.
How is authorization different from authentication, and why doesn't having a link prove someone is allowed to read the app?
In standard security terms, authentication establishes who is making a request, while authorization decides whether that requester may perform a particular action. A system can authenticate someone and still deny access to an object. It can also receive an anonymous request and decide whether that request is allowed for a genuinely public object.
The article’s example policy check asks whether a subject may read a particular artifact. Its four parts are who is asking, what object they want, what action they want, and what the system decides. A link identifies the object or provides a route to it; by itself, it does not answer whether this requester is allowed to read it.
In Vector Build, the article says authorization on read was missing, so possession of the reference became the whole access-control system. Claroc found many app addresses opened without sign-in. The article recommends a per-object permission check and an error for unauthorized reads, rather than treating “Anyone with the link” or a hard-to-guess hostname as the lock.
Why should an app check permission to read each object even when its URL is hard to guess?
A difficult URL can make guessing harder, but it does not decide who is allowed to read the object. The article says four random words were genuinely unguessable, yet passive DNS services collected hostname lookups into searchable datasets. Once an address is discovered or shared, secrecy of the URL no longer protects the data.
The proposed check is made during the read: fetch the object, then ask whether the subject may perform the read action. The article notes that the object has already been fetched to serve the response, so the permission decision can be another read against a policy store on the same request path. This is different from relying on a random identifier to hide the object.
The article says random IDs move the cost of finding an object from zero to an afternoon of scraping, while a per-object check uses a lookup already happening on the request path. It recommends checking every reachable surface and alerting on anonymous denials, since enumeration can arrive as a burst.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web