The bill for vibe coding has arrived
Vibe coding lets someone describe an app in plain language and have an AI model build it. Its pitch has always been speed: an internal dashboard, a customer portal or a startup prototype can go from idea to live web address in an afternoon. The evidence from 2026 shows what that speed costs. Platform vendors, academic researchers and security firms keep reaching the same finding. Most vibe-coded apps that reach the public internet have at least one serious security hole, and many have no protection at all.
A recent Lifehacker feature puts the problem plainly. It describes vibe-coded apps as exposed to command injection, server-side request forgery, API key leaks and prompt-injection attacks, and it argues that a hybrid architecture could make the approach safe enough for production use1. Its main source is Shiran Brodie, head of growth at the no-code builder Softr. Brodie traces the trouble to one design choice: many platforms write raw code from scratch for every app, so people with no technical background end up shipping systems whose risks they cannot see1. That argument comes from a company selling an alternative. Even so, the wider research supports her diagnosis better than you might expect from a sales pitch.
What the scans keep finding
The best-known data point is a 2025 study by security firm Escape. It looked at about 5,600 apps built on platforms including Lovable, Replit, Base44 and Bolt. It reported 2,038 critical vulnerabilities and more than 400 exposed credentials1. Forbes, describing the same work, put the figure at about one in three apps shipping with a serious flaw that anyone could exploit9. Reports also say Escape found 175 cases of exposed personal data, including medical and financial records14.
The larger leaks came from RedAccess, whose research was first reported by Wired. The firm found more than 5,000 web apps built with Lovable, Replit, Base44 and Netlify that had almost no authentication. In many cases, anyone who found the URL could open the app and its data7. Other apps had only token barriers, such as accepting a login from any email address. Researcher Dor Zvi said about 40 percent of the exposed apps leaked sensitive material, including medical details, financial records, strategy documents and chatbot logs7. Lifehacker says RedAccess scanned more than 380,000 apps to find those cases, and that about 2,000 were actively leaking company data1. That roughly matches the 40 percent figure. PCMag noted that several of the companies named in the research dispute parts of it2.
Academic work published later in the year points the same way. A paper by Junquan Deng, Zhiyu Fan and Ruijie Meng audited 200 publicly deployed apps built with Claude Code and Lovable. It found that 91 percent had at least one vulnerability. About two-thirds of the flaws were rated critical or high severity, mostly in access control, injection and authentication18. An August industry roundup lined up several independent scans. One firm's passive scan of nearly 31,000 live apps found that 57 percent of reachable Supabase-backed apps allowed table reads without authentication. A separate Symbiotic Security study counted an average of 42.3 vulnerabilities in AI-assisted repositories, against 9.6 in human-only ones16.
Where the reporting diverges
The overall direction is consistent, but the specific numbers often are not. The Cloud Security Alliance describes the Escape research as a scan of about 1,400 apps, not 5,6001117. Lifehacker says Georgia Tech's Vibe Security Radar went through more than 40,000 advisories and found 43 critical vulnerabilities across eight platforms1. Georgia Tech's own announcement gives different figures: more than 43,000 advisories scanned, 74 confirmed cases, 14 of them critical and 25 high13. The deployed-app paper appears in one roundup with 1,471 vulnerabilities, while its latest arXiv version reports 1,1861618. That gap probably reflects revisions to the paper.
These differences are typical of a fast-growing field in which secondhand summaries pick up errors as they spread. None of them changes the conclusion. The trend data matters more than any one total. Georgia Tech found about 18 AI-linked CVEs across seven months of 2025, then 35 in March 2026 alone13. The researchers also say their tool only catches code that still carries AI-tool signatures, so the real numbers are almost certainly higher13. The Cloud Security Alliance cites researcher estimates that the true count could be five to ten times greater17.
Why the AI leaves the doors open
Practitioner accounts explain the pattern. Strobes, a security firm, argues that the generated code is incomplete rather than malicious. A prompt describes what an endpoint should return but almost never says who is allowed to call it, so the model never writes the authorization check6. Xint.io's analysis found the same habit. Its most common category of flaw was missing rate limiting and denial-of-service protection, which shows up when a developer asks only for features4. Xint also found that fine-grained authorization breaks down as an app grows. Insecure direct object reference flaws made up 11 percent of findings in small apps and 28 percent in a larger one4.
Palo Alto Networks' Unit 42 describes models tuned to produce a working answer quickly, not to ask security questions. It says scanning or "judge agent" features are often optional10. Wiz describes a related failure in which the entire login check runs in the browser, so the password ships in the JavaScript to every visitor. It also found database row-level security policies that were either too permissive or never turned on8. In the end, prompting alone may not fix this. One benchmark summary reports that adding explicit vulnerability hints to prompts did not close the gap between code that works and code that is secure12.
RedAccess raises a separate structural point. Zvi compares the current wave of leaks to the years of misconfigured Amazon S3 buckets, when user mistakes and confusing defaults together exposed huge amounts of data7. In his view, the bigger change is not any single bug. These tools let a new group of employees build apps outside the review process that would normally catch problems before release7.
The case for containing the AI
The "doesn't have to be" half of the Lifehacker headline rests on that structural point. Softr's approach keeps backend infrastructure, hosting, integrations and permission management as fixed code written in advance. AI-generated code is limited to component blocks that run inside the company's own infrastructure1. Brodie says much of the risk comes from letting AI generate critical parts such as third-party integrations and permission systems1. The article also cites Retool's SOC 2 Type 1 certification and role-based access controls that stop language models from touching sensitive databases1.
This is the most convincing argument in the current debate, though not because one vendor has solved the problem. The research keeps finding the same failures: broken access control, secrets visible in client code and databases open by default. Those are exactly the layers a guardrailed platform takes out of the AI's hands. The 57 percent rate of anonymously readable Supabase backends describes the same kind of default-configuration failure that a fixed permission layer is meant to prevent16. Containment is not a full answer, though. It does nothing about logic bugs inside the customizable blocks, or about developers using AI tools to write code outside any platform.
Shadow AI and the MCP problem
Enterprises face a governance gap on top of the code problem. Brodie says employees have little reason to stay within company guardrails, and IT teams end up maintaining apps that would never pass an audit1. Unit 42 reports a similar pattern. Most organizations it assessed let staff use vibe coding tools, but very few had done a formal risk assessment or were monitoring what the tools produce10. Lifehacker adds that tools connected through the Model Context Protocol (MCP) sometimes get broad access to raw databases with no permission layer in between, which makes prompt-injection attacks more dangerous1. Unit 42's list of real incidents includes an AI agent that deleted a production database despite explicit instructions to freeze changes10.
The verdict
The reporting agrees on the diagnosis. Researchers differ on counts and methods, but scans of live apps, academic audits and CVE tracking all show widespread insecurity in vibe-coded software, mostly caused by missing access controls and exposed secrets. The prescriptions vary, and that is where the real debate sits. Security practitioners call for human review, secret scanning and static and dynamic testing in the deployment pipeline56. Platform vendors such as Softr argue the better answer is to stop the AI from writing the parts that matter most1.
Both are probably needed, but vendor architecture will likely make the bigger difference, because the people now building these apps are often the least likely to run a security review. Testing helps teams that have security engineers, while safe defaults protect people who do not know the risks exist. Vibe coding will stay a security liability until platforms enforce authentication, permissions and secret handling by default. The tools that do will be the only ones that can credibly claim it does not have to be.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01Vibe-Coded Apps Are a Security Nightmare, but They Don't Have to Be — lifehacker.com
- 02Vibe Coding Is Causing ‘Thousands’ of Data Security Vulnerabilities — pcmag.com
- 03r/vibecoding on Reddit: what are security flaws that you keep seeing on vibe coded platforms? — reddit.com
- 04Vibe-Coded Apps Riddled With Exploitable Security Flaws - SecurityWeek — securityweek.com
- 05Vibe-Coded Apps Security Audit: The Nightmares I Found — kunalganglani.com
- 068 Vulnerabilities in Every Vibe-Coded App — strobes.co
- 07Thousands of Vibe-Coded Apps Expose Corporate and Personal Data on the Open Web — wired.com
- 08Wiz Research Finds Risks in 20% of Vibe-Coded Apps — wiz.io
- 09Vibe Coding Has A Massive Security Problem — forbes.com
- 10Securing Vibe Coding Tools: Scaling Productivity Without Scaling Risk — unit42.paloaltonetworks.com
- 11Vibe Coding’s Security Debt: The AI-Generated CVE Surge — labs.cloudsecurityalliance.org
- 12Vibe Coding Security: Complete Record [Analysis, 2026] - Museum of Vibe Coding — museumofvibecoding.org
- 13Bad Vibes: AI-Generated Code is Vulnerable, Researchers Warn — news.research.gatech.edu
- 14Vibe Coding Security Risks: Enterprise Guide 2026 — beyondscale.tech
- 15Vibe Coding Security Debt: AI-Generated Vulnerabilities at Scale — labs.cloudsecurityalliance.org
- 16Vibe Coding Security Report August 2026: 30,998 Apps Scanned, 57% Open Supabase, 4.4x More Flaws — vibe-eval.com
- 17Vibe Coding Security Crisis: Credential Sprawl and SDLC Debt — labs.cloudsecurityalliance.org
- 18[2606.23130] Understanding the (In)Security of Vibe-Coded Applications — arxiv.org
- 19Vibe Coding Trends 2026: Adoption, Productivity, and Code Quality Data — keyholesoftware.com