AI Apps Expose Data via Open Defaults

AI-Built Apps Are Shipping With the Doors Open

The fastest way to leak company data in 2026 is to build something useful with an AI tool and forget to ask who else can open it. Two stories from the past week show the same failure from opposite directions. In one, employees created apps with an AI builder and left them reachable by anyone with a link. In the other, a hosting company’s connector for AI assistants became the way in.

Security researchers at Claroc found tens of thousands of apps built with Grok that opened without a sign in, including 24 built by employees of SpaceXSI and X. Separately, German web host Mittwald said attackers used a flaw in its Model Context Protocol server to compromise 56 customer projects. One story is about what people build with AI. The other is about what AI is allowed to connect to.

Both end in the same place. Sensitive content sits behind a door that nobody checked, and the organization finds out from someone else. For a CISO or Chief Compliance Officer, the lesson is not that one vendor failed. It is that AI lowered the cost of creating and connecting systems far faster than it lowered the cost of governing them.

Key Takeaways

  1. Easy to build means easy to expose. Claroc reported that 35,578 Grok built apps opened without an account, about 94.7 percent of the apps that answered.
  2. The builders were not careless outsiders. Twenty four of the open apps were created by SpaceXSI and X staff, which suggests default settings and habits matter more than expertise.
  3. AI connectors are an attack surface. Mittwald traced a breach of 56 customer projects to a flaw in the server that links its hosting panel to AI assistants.
  4. Detection lagged in both cases. Mittwald took eight days to notice exploitation, and the exposed Grok apps were found by outside researchers.
  5. Governance must start at the data. Controlling what AI built apps and AI connectors can reach, and logging it, matters more than policing every tool.

What Claroc Found in Grok Built Apps

Claroc published its findings in a report titled GhostShares 2. The numbers below come from that report, which describes its own methodology.

The researchers listed names under the grok.me domain from a commercial passive DNS service, pulling the list on 8 November 2026 after daily queries that began on 31 October. They counted 43,551 distinct apps that loaded, and 35,578 of them opened without sign in. Another 1,993 returned an authentication error. Claroc says roughly 94.7 percent of the apps that answered were open to anyone. The report does not account for the remaining apps in the total.

Then the content. Claroc’s AI assisted grading found that about one in 10 apps held sensitive material, internal business data or personal data. Roughly 2 percent were estimated to hold material rated high or critical, and the authors say their stricter second pass likely over counts, putting the real figure at roughly 650 to 800 apps. They confirmed 150 exposures by hand. Examples included an ICU patient briefing, a hospital staff roster, a fire department shift board and a timesheet app with plain text passwords. Thirty organizations were identified apart from SpaceXSI and X, 26 of them with more than 1,000 employees and six in the Fortune 500.

The method deserves a note, because it shapes how you read the numbers. Claroc requested each first page once without an account, did not use credentials it found, and rendered pages offline. A spot check of 100 AI graded apps found 75 held up. In other words, these are careful estimates, not an audit of every app.

SpaceXSI’s response, as the report relays it, was that Grok apps are private until the creator chooses to share them, and that employee accounts default to a setting called Only you. The researchers dispute that extra publishing steps protect data, noting that the option Anyone with the link was preselected in their own account. SpaceXSI took down the 24 staff built apps after a reporter’s questions, but not apps belonging to other users.

Defaults Are the Real Vulnerability

It is tempting to treat this as one vendor’s settings problem. That reading misses the pattern.

When a tool makes building trivial, the person building is rarely a security engineer. A marketing analyst makes a dashboard, a nurse manager makes a shift board, and a finance employee makes a timesheet app. Each wants to share it with colleagues, and the easiest way to share is a link. Whether the link requires a login depends on a setting most builders never see.

This is shadow IT with a faster clock. A decade ago the unsanctioned system was a spreadsheet emailed around. Today it is a hosted application with a public address, a database behind it, and possibly credentials embedded in it. The scale is different, and so is the discovery path. Attackers and researchers can enumerate public apps through DNS records without touching your network. If you wait for your own scanners to find them, you will be later than they are.

The NCA and CybSafe Oh, Behave! research reported that 32 percent of workplace AI users had shared sensitive company data without their employer’s knowledge. Pair that with an AI builder that publishes by default and you have the conditions for what Claroc found. Treat it as one survey’s finding, but notice how well it matches the behavior.

Mittwald and the AI Connector Problem

The second story involves software that sits between AI assistants and a platform. Mittwald CM Service, a German web host that serves agencies, runs a hosting management panel called mStudio. It offered a Model Context Protocol server so customers could connect AI assistants to that panel.

Mittwald’s own account, as reported by webhosting.today and The Record, runs as follows. An attacker first tested a vulnerability in the MCP server in October, moved through the platform as an ordinary user, and exploited the flaw on 27 October, five minutes after registering another account. The attacker read session tokens, not passwords, from customers who used the server, and compromised 56 projects. Mittwald noticed unusual behavior on 4 November, took parts of its systems offline, identified the MCP server as the entry point, and shut it down. It brought in external forensic specialists, informed the relevant authorities, and revoked tokens and blocked API keys as a precaution.

Several details are worth underlining. Mittwald says it has located and closed the flaw, and has not yet said whether the MCP server will return. Its current assessment is that customers who never used the server were not affected. It found no evidence so far that customer data was published, though checks continue. Its chief executive said he expects AI driven attacks to become more common, and the company said the attacker’s speed suggested AI tools were involved.

Mittwald says it located the flaw but does not describe it, so avoid claiming a cause beyond what the company has said.

What the Two Stories Share

On the surface these are unrelated. One involves a consumer facing AI builder, the other an enterprise hosting connector. Underneath, three things line up.

Authorization was loose. In the Grok case, an open link was the access control. In the Mittwald case, session tokens retrieved by the attacker were the key. Neither involved a sophisticated break. Both involved credentials or links that worked for anyone who held them.

Visibility was poor. Nobody at the affected organizations had a complete list of what had been built or connected. Mittwald took eight days between exploitation and detection. Claroc found exposures through public DNS data that any outsider could use.

The blast radius followed the data. Tokens opened projects, and open links exposed business content. What an attacker could reach depended on what the app or connector was allowed to reach, not on how clever the attack was.

This is why zero trust generative AI thinking helps. It treats every AI built app and every AI connector as an untrusted accessor. It asks for authentication on each request, limits what each can reach, and records what happened.

What a CISO Should Do This Quarter

Start by finding what exists. Use external asset discovery to enumerate public hostnames tied to your company, including those on AI builder domains, and ask business units what they have created. Do not rely on a policy telling people not to build things. Assume they did.

Next, set a rule for AI builders. Decide which tools are approved, require sign in for anything that touches business or personal data, and make private the default. Where a tool cannot enforce that, restrict it to non sensitive uses.

Then inventory your AI connectors. List every MCP server, plugin and API connection that lets an AI assistant act on your systems. For each, record who can use it, what it can read and write, how it authenticates, and where its logs go. The Mittwald case shows what happens when the connector is a trusted path with a flaw in it.

After that, shorten detection time. Feed connector and application logs into your SIEM, and alert on unusual token use, new accounts followed quickly by privileged actions, and access to sensitive paths. Eight days is a long time for an attacker to hold a session.

Finally, rotate credentials after any incident that involves tokens. Mittwald’s guidance to customers, to change passwords, enable two factor authentication, review API keys and renew third party credentials stored in projects, is a reasonable template. MFA will not stop a stolen session token by itself, but it narrows the paths around it.

What a Chief Compliance Officer Should Ask

Compliance leaders face a different version of the problem. An AI built app that holds personal data can create a privacy incident with no malicious actor. If the app contains health information, HIPAA duties follow. If it holds data on EU residents, GDPR obligations apply regardless of who built the app or how quickly.

Ask five questions. Which AI builders and connectors are in use, and who approved them? What categories of regulated data have been placed in them? Which of them require authentication? Where are their logs, and for how long are they kept? And what is the process for discovering, assessing and, if needed, reporting an exposure?

If the answers depend on goodwill rather than records, you have the same visibility gap Claroc and Mittwald exposed. The evidence a regulator or auditor will ask for is an inventory and a log, not a statement of intent. Strong access controls and third party risk management practices apply to AI vendors as they do to any supplier.

Governing the Data Behind the Apps

Policing every AI tool is a losing race. New builders and connectors appear weekly. The durable control sits at the data, where policy can follow content regardless of which tool asks for it.

The pattern has four parts. AI assistants and agents reach sensitive content only through a controlled channel. Each request is authenticated and evaluated against attribute based policies. Authentication credentials stay outside the model’s reach, so a prompt cannot extract them. And every operation lands in one audit log that the security team can query.

Kiteworks builds this into its Secure MCP Server, which lets AI assistants interact with content in the Kiteworks secure data exchange while inheriting the user’s permissions, with tokens kept in the operating system’s credential store. It also provides Kiteworks Compliant AI, which extends the same policy engine and logging to AI workflows. The honest scope is narrow. Kiteworks does not secure apps built in other AI tools, and it does not run inside a hosting provider’s control panel. It governs AI access to the content held in the Kiteworks platform and records that access. For an organization whose most sensitive files live there, that turns a pile of unknowns into a bounded question with a log behind it. For the broader design, see AI data protection and role based access control.

The Larger Lesson

AI changed who can create software, and how quickly. It did not change what attackers want, which is accessible data. Every organization will have more apps, more connectors and more people building than its security team can review by hand.

So decide in advance what will not be negotiable. Sensitive data is reachable only through governed channels. Every AI accessor is authenticated and limited. Every access is logged. Then measure how long it takes you to discover something you did not know existed. If the answer is measured in days, as at Mittwald, or never, as with the open Grok apps, that is the number to improve first.

Why Traditional Application Security Misses These Apps

Application security programs grew up around code that engineering teams write, review and deploy through controlled pipelines. The apps in the Claroc research do not follow that path. They are created by business users inside a vendor’s environment and published with a click.

That breaks several assumptions. There is no pull request to review, no scan in a build pipeline and no ticket in the security queue. The hostname may sit on the vendor’s domain, so it does not appear in your DNS inventory. The data may enter through a form the builder designed, with no data classification applied. And the person responsible may not think of themselves as having built software at all.

A control strategy must start from where the exposure occurs. Discovery should look outward, using external attack surface tools and passive DNS searches for your company names and brands on AI builder domains. Governance should look at data, since the dangerous part of an app is the content it holds. And policy should reach the user, since the decision to publish is theirs.

Consider a simple rule set. Any AI built app that touches customer, patient, employee or financial data must require authentication and be registered with a named owner. Apps without registration are removed after notice. Sensitive categories are blocked from AI builders altogether unless the builder is an approved enterprise tool with contractual protections. These rules are easy to state and easy to enforce once you can see the apps.

A Review Gate for Connectors and Builders

Every new AI builder, plugin or MCP connector should pass through a lightweight review before use with company data. The review does not need to be slow. A one page checklist can cover most of the risk.

Authentication. Does the tool support single sign on and require login for published output by default? Can administrators enforce that setting?

Authorization. What can the connector read and write? Can its permissions be scoped to a user’s own access, or does it run with a broad service account? In the Mittwald case, tokens retrieved by the attacker were the prize, so how tokens are issued, stored and revoked matters.

Logging. Does the tool record who used it, what it touched and when? Can those logs reach your SIEM in a usable format?

Data handling. Where does data go, how long does it stay and is it used for training? What happens when the project is deleted?

Vulnerability handling. How does the vendor disclose and patch issues? Is there a security contact, and what response times do they commit to?

Exit. Can you disable the tool quickly if something goes wrong, and can you export or delete your data?

Record the outcome and revisit it annually, or sooner if the vendor suffers an incident. This process mirrors what many teams already do for SaaS applications. The difference is volume and speed, so keep it light enough that teams use it. Complement it with third party risk management reviews for the larger vendors and with supply chain risk management practices for open source components.

Metrics That Show Whether It Is Working

Without measures, governance turns into a document. A handful of metrics keep it honest.

Count of known AI built apps and connectors, and share with a named owner. Share requiring authentication. Mean time from creation of an exposed app to its discovery by your own team, compared with when an outsider could have found it. Count of connectors with broad service account access. Mean time to revoke tokens after an incident. And share of AI data access that flows through a governed, logged channel.

Track these over time and present the trend. A rising share of governed access and a falling discovery time tell a better story than a static count of risks. If the numbers stall, investigate why, which often points to a business unit that finds the approved route too slow.

Be careful with vanity numbers. The number of blocked requests, for instance, can rise because attackers are probing or because your tooling is noisy. Prefer outcome measures tied to exposure and evidence. Pair them with periodic exercises. A red team that tries to find your own employees’ public AI built apps will teach you more than a dashboard.

To see how a governed AI connector enforces policy and logs every request, request a Kiteworks demo and walk through the audit trail with our team.

Frequently Asked Questions

Use external asset discovery to look for hostnames tied to your company on AI builder domains, search passive DNS data, and ask business units directly with no penalty for honesty. Combine results with network and identity logs. Then require sign in and approval for anything touching business data. See AI data governance.

A Model Context Protocol server lets AI assistants connect to a system and perform actions through a standard interface. It becomes a risk when it grants broad access or has flaws, as in the Mittwald case. Limit its permissions, authenticate every request and log everything.

It varies by tool and by settings. In the Claroc research, the option Anyone with the link was preselected in the researchers’ own account, though SpaceXSI said apps are private until the creator shares them. Check each tool’s defaults yourself and make private the standard. Strengthen access controls around them.

Take the connector offline, revoke tokens and API keys, review logs for activity from the exposure window, rotate credentials stored in affected systems, and notify affected parties as required. Mittwald’s response is a reasonable model. Follow your incident response plan.

Not on its own. MFA protects the login step, but a stolen session token can bypass it. Combine MFA with short token lifetimes, token revocation, conditional access and monitoring for unusual use. Learn more about MFA in context.

Additional Resources

Get started.

It’s easy to start ensuring regulatory compliance and effectively managing risk with Kiteworks. Join the thousands of organizations who are confident in how they exchange private data between people, machines, and systems. Get started today.

Share
Tweet
Share
Explore Kiteworks