Replit Security Under an Independent Pentest

An independent pentest by security firm Astra of an application built on Replit from a single prompt found gaps in rate limiting, security headers, and server-side access control, part of a study that produced 137 confirmed vulnerabilities across four AI-built applications. Replit received no advance notice of the test and has no vendor relationship with Astra, and Astra said the study did not rank Replit against Emergent, Lovable, and Base44 because all four apps shared the same nine weak points. Astra said the findings sit at the configuration and data layer rather than in application logic — "decisions nobody made, because nobody was asked to make them" — and recommended replacing vague prompt adjectives with named controls, such as locking an account after five failed attempts and returning a 429.

Key Takeaways - The Security Risks with Vibe Coding: A study by Astra uncovered 137 confirmed vulnerabilities across four AI-built applications. - Replit security testing revealed gaps in rate limiting, security headers, and server-side access control. - Functional AI-generated apps still need independent security testing before launch. - Astra’s autonomous pentesting helps identify and validate vulnerabilities in real applications before attackers do. We built an application on Replit from a single prompt, gave it no special instructions about security, and then pentested it the way an attacker would. Replit had no advance notice that any of this was happening, and there is no vendor relationship between us. The app worked, which is worth saying first, because nothing about the build looked wrong from the outside. It was also one of four AI-built applications in a study that produced 137 confirmed vulnerabilities https://www.getastra.com/reports/vibe-coding-security-risks-study between them, and this one carried its share. Is Replit safe to use? People usually mean one of two things by this, and the answer differs depending on which. The first question is whether Replit as a platform is secure, meaning its infrastructure, hosting, and account systems. That is a question about Replit the company, and it is not what we tested. The second is whether the application Replit’s agent writes for you is secure. That is what we tested, and it matters more if you are about to put something in front of real users. A working app and a secure app are different bars, and an AI coding tool is optimising for the first one. What we found in the Replit-built app Replit received the same prompt as Emergent, Lovable, and Base44, with the same wording and the same lack of warning. The prompt described the application we wanted built and said nothing about security requirements, which is how most people use these tools in practice. Every finding went through the same pipeline before it counted. Two autonomous pentest agents tested each app independently, one mapping the application and building attack chains, the other going in cold like an unbriefed attacker. A validator agent then reproduced every finding from scratch, and a human security engineer reviewed what was left. Anything nobody could independently confirm was cut. It is worth being precise about what this shows. The study did not rank the four platforms against each other, and we are not claiming Replit is more or less secure than the alternatives, because all four apps shared the same nine weak points. What it shows is how one Replit-built application held up under real testing. Replit agent security and why the gaps show up The agent builds what the prompt describes, and our prompt described an application rather than a security posture. What came back matched the request. Almost everything we found sits at the configuration or data layer rather than in the application logic. Rate limiting that was never set up. Headers that were never added. Authorization enforced in the interface but not on the server, which is the pattern behind most broken access control https://www.getastra.com/blog/api-security/broken-object-level-authorization-bola/ findings. These are less like bugs in the generated code and more like decisions nobody made, because nobody was asked to make them. That is the awkward part of vibe coding security. The tool is doing its job correctly, on a specification that was only half written. How to harden a Replit app before you ship Three things move the needle most, and none of them needs a security background. Start with the prompt, replacing every adjective with a number or a named control, because “secure login” gets you nothing while “lock the account after five failed attempts and return a 429” is a requirement the agent can implement. Replit’s own security documentation https://docs.replit.com/category/security is worth reading alongside this, since some protections are configured on the platform rather than written into your code. Enforce access control on the server and in the database, not only in the interface. Hiding a button is not authorization. Broken access control was consistently the most severe failure in the study whenever it appeared, because it lets one user reach another user’s data. Then test the running application. Try logging in eight times with the wrong password and see whether attempt six is rejected. Open DevTools and check whether a Content-Security-Policy https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Content-Security-Policy header exists at all. Reset a password with a fake email and then a real one, and watch for any difference in the response or its timing. Those three checks take about ten minutes and will tell you more than reading the code will. If you want the longer version, our API penetration testing https://www.getastra.com/blog/security-audit/api-penetration-testing/ guide covers what a full pass looks like. Final thoughts Replit gets you from an idea to a working application faster than almost anything else, and that is the point of it. Shipping to real users is a different bar, and clearing it means asking for things the prompt did not cover. Switching platforms will not fix it either. All four tools in this study produced the same nine weak points from the same prompt, which suggests the gap is in how we are asking rather than which tool we are asking. If you want this run against your own application instead of a demo build, that is what Astra’s autonomous pentest https://www.getastra.com/autonomous-pentesting does, using the validation pipeline described above. The full findings across all four apps, plus the hardened prompt, are in the free report https://www.getastra.com/reports/vibe-coding-security-cost-study .