Industries / Technology / Software and SaaS / Low-code platforms
Control and oversight
Keeping apps built outside the IT department safe, known and within the rules: who can see which data, who owns each app, and what it may do.
5
Milestones achieved, from 2021 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2021
Defaults exposed 38 million records
UpGuard, a security firm, found Power Apps portals, public websites built on Microsoft's platform, whose lists of records could be read by anyone. Lists ignored table permissions unless the builder switched them on. UpGuard notified 47 organizations, including American Airlines and the state of Indiana, about 38 million records in all. Microsoft then made table permissions on by default.
Lessonwhat a platform does by default is what most of its apps will do.
Credit
UpGuard
2022
Ten risks for citizen apps
The OWASP Foundation, which publishes free security guidance, released the Low-Code/No-Code Top 10, a list of the main security risks in apps built by business users, in July 2022. In November 2025 it renamed the list the Citizen Development Top 10 and widened it to AI-assisted coding and AI agents.
Credit
the OWASP project's volunteers; its leaders are now Kayla Underkoffler, Shauna Rathbun, Ziv Hagbi and Ory Segal
2023
Guest access left open
KrebsOnSecurity reported that banks, healthcare providers and the state of Vermont had Salesforce Community websites that let visitors who had not logged in read private records. One Vermont site exposed pandemic assistance applicants' Social Security and bank account numbers. The cause was administrators granting the "guest user" too much access.
Credit
researcher Charan Akiri and Brian Krebs; Aaron Costello described the flaw in 2021
2024
Chatbots public by default
At the Black Hat conference, Michael Bargury of Zenity, a security firm, said Copilot Studio, Microsoft's no-code tool for building chatbots, published bots to the web without requiring sign-in by default, and that his team had found tens of thousands online. Microsoft changed the default, but The Register reported the change applied only to new installations.
Credit
Zenity; The Register
2025
A form steers an agent
Noma Security found that instructions hidden in a Salesforce Web-to-Lead form, a public form that creates sales leads, could make Agentforce, Salesforce's AI agent, send customer records to an outside web address. The flaw was named ForcedLeak. Salesforce began enforcing lists of trusted web addresses on 8 September 2025.
Lessonan AI agent that reads text typed by strangers can be steered by that text.
Credit
Noma Security
Watch for
Five thousand exposed AI-built apps
RedAccess, an Israeli security firm, says it found 380,000 publicly reachable apps, databases and related services built with Lovable, Base44 and Replit or hosted on Netlify, and that about 5,000 (1.3%) held sensitive company data. Lovable and Base44 said RedAccess had not given them the web addresses needed to check, and Replit said it had 24 hours' notice.
Achieved whenthe platforms or an outside party confirm a published sample of the findings.
Claimed by
RedAccess, May 2026
One list of every agent
Microsoft took Agent 365 out of preview. It is a registry meant to give IT staff "a single view of every agent" in their organization, whether built with Copilot Studio, run on AWS Bedrock or installed by employees without approval, which Microsoft calls shadow AI.
Achieved whenan outside test plants a known set of agents in a test organization and publishes how many Agent 365 finds.
Claimed by
Microsoft, May 2026
Open
01
Default exposure, visual builders
A test, repeated every quarter, of what six visual builders expose on their default settings: Power Pages (formerly Power Apps portals), AppSheet, Bubble, Glide, Softr and Airtable interfaces. Build the same small app in each, with user accounts, records private to each user and an admin page. Then try, from a logged-out browser and from a second ordinary account, to read or change the first user's records, including by sending requests straight to the app's data services. Test only apps you built.
The briefCredit · unclaimed02
A checker for Power Platform
An open-source tool that reads an exported Power Platform solution, the zip file that holds a set of apps, flows and Copilot Studio agents, and flags the risks on OWASP's Citizen Development Top 10. Examples: tables that a public Power Pages site lets anyone read, connections shared with everyone, flows that send data to outside web addresses, passwords typed into flows, and agents that allow chat without sign-in. It writes its results in SARIF, the format GitHub's code scanning reads.
The briefCredit · unclaimed
Open 01 · the brief
Default exposure, visual builders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test, repeated every quarter, of what six visual builders expose on their default settings: Power Pages (formerly Power Apps portals), AppSheet, Bubble, Glide, Softr and Airtable interfaces. Build the same small app in each, with user accounts, records private to each user and an admin page. Then try, from a logged-out browser and from a second ordinary account, to read or change the first user's records, including by sending requests straight to the app's data services. Test only apps you built.
Why it is still open
Controlled tests on apps the testers built themselves have covered AI app builders: VibeTierList's second test (October 2026) built one team app three times each on Lovable, Bolt and Base44. The 2021 Power Apps finding came from scanning other organizations' live sites. A Cloud Security Alliance research note (November 2026) says that apps being public by default "is not universal, since at least one major builder already defaults new apps to private," and cites no comparison. No test found puts the visual builders under one written specification.
Built with what exists
free or trial accounts; Playwright to script both users; mitmproxy or Burp Suite Community Edition to capture and replay the app's requests; the OWASP Citizen Development Top 10 as the checklist.
First step
Write the specification and its 20 checks, and run them on a Glide app and a Bubble app.
Done looks like
6 builders, each app built twice from the same written specification, 20 access checks per app, 240 results per quarter, published with the specification and the scripts. Each maker gets 30 days' notice of any failure before publication.
Open 02 · the brief
A checker for Power Platform
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open-source tool that reads an exported Power Platform solution, the zip file that holds a set of apps, flows and Copilot Studio agents, and flags the risks on OWASP's Citizen Development Top 10. Examples: tables that a public Power Pages site lets anyone read, connections shared with everyone, flows that send data to outside web addresses, passwords typed into flows, and agents that allow chat without sign-in. It writes its results in SARIF, the format GitHub's code scanning reads.
Why it is still open
Microsoft's own solution checker has two rules in its security category, "avoid-eval" and "desktopflow-avoid-unsafe-password", and none for sharing, table permissions or connections. The open-source scanners found, such as lcnc-security-engine, read n8n and Make workflows, not Power Platform solutions, and that one implements three of the ten OWASP risks (CD-SEC-01, 04 and 05).
Built with what exists
a free Power Platform developer environment; the Power Platform command-line tool (pac) to export and unpack solutions; Python or Go; the OWASP list as the rule set; GitHub code scanning to display results.
First step
Export a solution containing a Power Pages site whose table permissions are off, unpack it, and write the first rule that finds the problem.
Done looks like
at least one rule for each of the ten risks. On a published set of 50 solutions with 100 planted flaws, it finds at least 90 with no more than 10 false alarms. A second person plants 20 new flaws without telling the author, and the catch rate on those is published.