← Wednesday, November 18 deck

64Security1 source posting

Viral Post Warns That Blind Vibe Coding Is Dangerous, Sharing a Builder's Account of Leaked API Keys

This is exactly why blind vibe coding is dangerous. 馃拃

Screenshot of a Reddit post titled I checked my own vibe coded app before launch, describing secret keys found in the app's files.Second screenshot of the Reddit post, describing how the author rotated the keys and capped spending.

Context

Why it matters

The story shows a common risk for people building apps with AI tools: secret keys shipped to every visitor's browser, with bots already probing for them within days.

Mashood K posted that this is exactly why blind vibe coding is dangerous, sharing screenshots of a Reddit post. In it, a non developer says they built a booking app with Cursor and Claude, and a friend found the OpenAI key, the live Stripe secret key and an email service key in plain text in files every visitor downloads. The author says server logs showed 212 requests for the app's environment file in two weeks, and that they rotated all three keys and added hard spending caps.

Deep dive

Mashood K posted that this is exactly why blind vibe coding is dangerous, with two screenshots of a Reddit post from the TechLeader community. The account in the screenshots is the Reddit author's, not a verified incident report, so it is attributed to them.

The author says they are not a developer and built a booking app for a partner's studio with Cursor and Claude, using Supabase, Stripe, OpenAI for a chat helper, and Resend for email. Before launch, a friend who works on backends opened the browser's network tools and searched the app's JavaScript files. According to the post, the OpenAI key was there, as was the live Stripe secret key and the Resend key, in plain text, in files every visitor downloads.

The author says they thought a hidden environment file kept the keys safe, but any variable starting with the VITE prefix is put into the frontend on purpose, and Claude had named the keys that way so the chat would work from the browser. Server logs showed 212 requests for the environment file in the two weeks the app was online, which the author attributes to bots that try every site. The OpenAI key had no spending limit.

The author says they rotated all three keys, put a 20 dollar hard cap on OpenAI, moved the OpenAI and Stripe calls into Supabase edge functions so the browser never touches a key, and stored keys in Infisical. Their advice is to search a build's output folder for secret key prefixes and to set hard spending caps.

What to expect next

Expect more tools that scan AI generated apps for exposed secrets before deployment, and more default warnings from coding assistants when keys are placed in frontend code. Watch for Cursor, Anthropic, Vercel and others to add guardrails for this exact pattern. The story is a reminder for new builders to check build output and set spending caps before launch, not after a bill arrives.