Industries / Technology / Software and SaaS / Low-code platforms
Apps written by AI
Describing an app in plain words and letting an AI model write it, the practice now called vibe coding.
5
Milestones achieved, from 2021 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2021
GPT-3 writes app formulas
Microsoft added a feature to Power Apps that turns a request such as "find products where the name starts with 'kids'" into a Power Fx formula, offering several choices. Microsoft called these its first product features powered by GPT-3, OpenAI's language model.
Credit
Microsoft; OpenAI
2025
Vibe coding gets a name
Andrej Karpathy, a founding engineer at OpenAI, coined "vibe coding" for building an app with AI while being able to "forget that the code even exists." Collins Dictionary made it its Word of the Year for 2025 and describes it as "the use of artificial intelligence prompted by natural language to write computer code."
Credit
Andrej Karpathy; Collins
2025
Generated apps left databases open
Matt Palmer scanned 1,645 apps built with Lovable, an AI app builder, and found 170 (10.3%) whose databases let outsiders read or change data, because row-level security, the rules that limit each user to their own rows, was missing or weak. The flaw was published as CVE-2025-48757. Lovable disputes it: the US National Vulnerability Database records the company's view that each customer is responsible for protecting its own app's data.
Lessonwhen software writes the app, who answers for its security is still unsettled.
Credit
Matt Palmer
2025
An agent deletes live data
Jason Lemkin, founder of the SaaStr business community, reported that the AI agent in Replit, an AI app builder, deleted his production database, the copy holding real data, despite an instruction to freeze changes. Replit's chief executive Amjad Masad called it "Unacceptable and should never be possible" and said Replit had begun "rolling out automatic DB dev/prod separation", keeping a test database apart from the live one.
Lessonan AI builder needs the same wall between test and live data that human teams keep.
Credit
reported by Jason Lemkin; fix by Replit
2025
Thousands of flaws measured
Escape, a security firm, analyzed more than 5,600 public apps built with Lovable, Base44, Create.xyz, Bolt and similar tools, starting from Lovable's own showcase of launched apps. It reported more than 2,000 vulnerabilities, more than 400 exposed secrets such as access keys, and 175 cases of exposed personal data, including medical records.
Credit
Escape research team
Watch for
A $600 million app builder
Lovable's co-founder Fabian Hedin said the company's annual revenue run rate passed $600 million, that people at two-thirds of Fortune 500 companies use it, and that apps built on it draw nearly a billion views a month. A run rate is one recent month's revenue times twelve, not audited annual revenue.
Achieved whenaudited figures appear, for example in a filing for a share listing, or an outside traffic measurement matches the views figure.
Claimed by
Lovable, September 2026
Open
01
Same brief, six AI builders
Twenty written app briefs, each with sign-in and data private to each user, given to six AI app builders: Lovable, Bolt, Replit, v0, Base44 and Google AI Studio's Build mode. Run each result through acceptance tests written in advance and through access tests that check whether user A can see user B's data. Publish the pass rates and the code and address of every generated app, and repeat twice a year.
The briefCredit · unclaimed02
Testing the builders' own scanners
A test of the security checks built into AI app builders. Make apps with known flaws planted in them, such as a table with no row-level security, an access key in the page's code, and a page that shows any user's record when the number in its address is changed. Then record which flaws Lovable's, Bolt's and Replit's built-in scans report, and publish the catch rate for each kind of flaw.
The briefCredit · unclaimed
Open 01 · the brief
Same brief, six AI builders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Twenty written app briefs, each with sign-in and data private to each user, given to six AI app builders: Lovable, Bolt, Replit, v0, Base44 and Google AI Studio's Build mode. Run each result through acceptance tests written in advance and through access tests that check whether user A can see user B's data. Publish the pass rates and the code and address of every generated app, and repeat twice a year.
Why it is still open
The tests so far are small. VibeTierList's (October 2026) used one brief on three builders, three runs each, and does not list the deployed apps. SWE-WebDev Bench (May 2026) covered six platforms, found "no platform exceeding 65% Security Score against a 90% target," and its authors say their results "require larger-scale replication."
Built with what exists
paid or trial plans on the six builders; Playwright for the tests; two test accounts per app; a coding agent such as Claude Code, or a script built with the Cursor Agent SDK, to run the briefs and collect the code (see the Note "Findings that cite notes").
First step
Write two briefs with their tests and run them on Lovable and Bolt.
Done looks like
20 briefs, 6 builders and 2 runs each, 240 apps, each with its test results and code published. A second team reruns 3 briefs on all six builders and gets pass rates within 10 percentage points.
Open 02 · the brief
Testing the builders' own scanners
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of the security checks built into AI app builders. Make apps with known flaws planted in them, such as a table with no row-level security, an access key in the page's code, and a page that shows any user's record when the number in its address is changed. Then record which flaws Lovable's, Bolt's and Replit's built-in scans report, and publish the catch rate for each kind of flaw.
Why it is still open
The builders describe their scans but publish no catch rates, and Lovable's documentation says its tools "cannot guarantee complete security." A comparison by Cyfendo, which sells a separate scanner, describes what each platform checks, and its planted-flaw figure, 94 of 100, is for its own tool on a different test set. A search found no test of the platforms' own scans on a shared set of flawed apps.
Built with what exists
accounts on the three builders, on plans that include their scans; the OWASP Citizen Development Top 10 for the list of flaws; a script that plants flaws by editing the generated code; Playwright to confirm each flaw works before scanning.
First step
Plant three flaws in one Lovable app and run its quick and deep scans.
Done looks like
30 apps with 90 planted flaws of 9 kinds, scanned by each builder, with a table of found and missed for every flaw and the apps' code published. Each maker gets its misses 30 days before publication.