Base44 Vibe Coding Flaw: How a Logic Bug Exposed Private Apps
In July 2025, Wiz Research found that Base44, the Wix-owned AI app-building ("vibe coding") platform, exposed undocumented registration and email verification endpoints. Anyone who knew an app's app_id, which is visible in the app's URL, could register for a private app and bypass its authentication, including single sign-on.
Topic: AI in Cybersecurity in Cybersecurity & Ethical Hacking · News & analysis
Also covers: Cyber Threats, Scams & Online Safety
Quick answer: In July 2025, Wiz Research found that Base44, the Wix-owned AI app-building ("vibe coding") platform, exposed undocumented registration and email verification endpoints. Anyone who knew an app's app_id, which is visible in the app's URL, could register for a private app and bypass its authentication, including single sign-on. Wiz disclosed it on 9 July 2025, Wix fixed it in under 24 hours, and Wix reported no evidence of past abuse. The lesson: hidden endpoints and non-secret identifiers are not access control.
Key takeaways
- The flaw was in the platform's authentication logic, not in users' generated code.
- An identifier visible in URLs was enough to register for a private app.
- The fix was platform-wide, so app owners did not need to act.
- Logic flaws are rarely found by scanners; test authorisation paths manually.
- Apps built with AI tools still need a security review before launch.
Table of Contents
- What Happened?
- What Was the Timeline?
- Why Is This a Logic Flaw?
- How Do You Test for Similar Flaws?
- What Should Teams Using AI Builders Learn?
- What Should a Review of an AI-Built App Cover?
- What Does a Correct Server-Side Check Look Like?
- References
What Happened?
According to Wiz's write-up, two authentication-related API endpoints were reachable without restriction. By supplying only an app's app_id, a researcher could create a verified account on an app configured as private, then log in, bypassing the owner's access settings and SSO.
What Was the Timeline?
| Date | Event |
|---|---|
| 9 July 2025 | Wiz reports the issue to Wix |
| Within 24 hours | Wix deploys a platform-wide fix |
| Late July 2025 | Wiz publishes details; Wix states no evidence of abuse was found |
The Hacker News covered the disclosure at the time.
Why Is This a Logic Flaw?
Each endpoint worked as coded. The failure was in the rules: registration should have checked whether the app allowed self-registration. This is broken authentication as described in OWASP API Security Top 10, API2. Automated scanners see valid responses and move on; only someone asking whether a stranger should be able to do this finds it.
How Do You Test for Similar Flaws?
On your own app, test every route with an unprivileged account and no account. A small authorisation test with Python:
import requests
BASE = "https://staging.example.com/api"
routes = ["/apps/123/users", "/apps/123/data", "/auth/register"]
sessions = {
"anonymous": requests.Session(),
"outsider": requests.Session(), # logged in to a different app or tenant
}
sessions["outsider"].headers["Authorization"] = "Bearer OUTSIDER_TEST_TOKEN"
for name, s in sessions.items():
for r in routes:
code = s.get(BASE + r, timeout=5).status_code
flag = "CHECK" if code < 400 else "ok"
print(f"{name:10} {r:22} {code} {flag}")
Any success code for a role that should be denied needs investigation. Also list endpoints from the JavaScript bundle and API documentation, since undocumented ones are often the weakest.
What Should Teams Using AI Builders Learn?
- Treat identifiers in URLs as public; never rely on them as secrets.
- Ask the platform how private access is enforced and how incidents are reported.
- Keep sensitive data out of AI-built apps until a security review is done.
- Add access tests for every role to your release checklist.
- Watch the platform's security notices and apply any required actions.
Read our guides on prompt injection and penetration testing techniques, or talk to a WebAsha training advisor about web application security.
What Should a Review of an AI-Built App Cover?
The OWASP Application Security Verification Standard (ASVS) and the Web Security Testing Guide, including its business logic tests, give a structure. A short review for an app built on an AI platform:
| Area | Question to answer | How to check |
|---|---|---|
| Authentication | Can someone register or log in where they should not? | Try sign-up on a private app; test every auth endpoint |
| Authorisation | Can user A read or change user B's records? | Swap record IDs between two test accounts |
| Data exposure | Do API responses include fields the screen does not show? | Inspect responses in browser developer tools |
| Secrets | Are API keys in the front-end bundle? | Search the built JavaScript for key patterns |
| Business logic | Can steps be skipped, such as payment before order confirmation? | Replay requests out of order |
| Third-party services | Which external APIs receive user data? | Review integrations and their terms |
# look for exposed keys in a built front-end
grep -RoE "(sk_live_[0-9a-zA-Z]{20,}|AIza[0-9A-Za-z_-]{35}|AKIA[0-9A-Z]{16})" dist/ | sort -u
What Does a Correct Server-Side Check Look Like?
The Base44 issue came down to a registration path that did not ask whether the app allowed outsiders. In your own code, every sensitive route should check the app's rules and the user's rights on the server. A minimal Flask example:
from flask import Flask, abort, g, request
app = Flask(__name__)
APPS = {"123": {"private": True, "allowed_domains": {"example.com"}}}
@app.post("/apps/<app_id>/register")
def register(app_id):
cfg = APPS.get(app_id) or abort(404)
email = request.json.get("email", "")
domain = email.rsplit("@", 1)[-1].lower()
if cfg["private"] and domain not in cfg["allowed_domains"]:
abort(403) # a known app_id alone must never grant access
# ... create the pending account and send verification
return {"status": "verification sent"}, 202
The rule is enforced where the attacker cannot change it, and the identifier in the URL only selects which rules apply.
References
Frequently Asked Questions
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0