The Vibe Coding Debate Just Got Its Answer

Moltbook, Lovable, and the Tea app didn’t settle who’s right. They settled who pays first.

Photo by AltumCode on Unsplash

I’m not a technical person, and I don’t pretend to follow everything about AI or how software actually gets built. But I read a lot. I keep running into articles and discussions on Medium and other tech platforms, and every so often one of them stops me mid-scroll. That’s honestly how I ended up writing about AI in the first place. **In an earlier piece**, I laid out my own guess about where this was heading, and at the time I thought some of it might take five years, maybe longer, to really show up. Looking at things now, I keep going back to that piece and asking myself whether I got it right, or whether I just didn’t see how fast this would move.

Here’s what’s happened since. I’d written that the real stress test hadn’t even begun. It has now. And honestly, it’s still going — new incidents keep surfacing even as I’m writing this.

Moltbook, and the founder who skipped the code entirely

Back in February, a social network built entirely for AI agents — it’s called Moltbook — ended up in international security news. The person behind it made no secret of how it came together: rather than sitting down and coding it himself, he described what he wanted and let an AI system build the whole thing.

Within about 72 hours of launch, the platform had exposed roughly 1.5 million API tokens and tens of thousands of email addresses. When researchers looked into the cause, there was nothing clever about it. A database had been left completely open — readable and writable by anyone who found it. Nobody had made a deliberate choice to leave it that way. It seems more that nobody made any choice at all; the AI defaulted to open access, and no one circled back to close the door.

That’s the part that stays with me. Calling it a hack gives it too much credit. It was really just a gap nobody noticed, because nobody thought to look.

Lovable, and a company still working out what “private” means

Then there’s Lovable — a vibe-coding platform worth billions, used by millions of people to build apps just by describing what they want. Over roughly two months this year, it went through three separate security incidents, each one exposing source code, database credentials, user records.

What got me wasn’t just the breaches themselves. It was watching the company’s account of what happened keep shifting — first that this wasn’t really a breach at all, then that the exposed data had been meant to be visible, then pointing at unclear documentation, then at a bug bounty partner who’d handled the report. Every version moved the responsibility somewhere else. And that tells you something the breach on its own doesn’t: even the companies building these tools haven’t settled, internally, what “private” is supposed to mean inside the systems they’ve built.

The Tea app, and a warning we’d already been given

Most people have probably already heard about the Tea app breach in some form. Private messages between users ended up visible to strangers, not through some elaborate exploit but through access-control logic that nobody seems to have reviewed for the obvious question — who’s actually allowed to see this?

I want to be careful here about what’s confirmed and what isn’t. Tea’s breach happened back in July 2025, months before Moltbook or Lovable made headlines, and whether it was actually built through vibe coding is still genuinely disputed. Some of the researchers who looked closely think it’s more likely an ordinary beginner mistake than proof of AI-generated code specifically. What isn’t in question is the mistake itself, or the fact that it happened earlier than most people assume. If anything, that timing makes this feel less like a brand-new problem and more like something we’d already been warned about, before anyone had a name for it.

It’s not a new category of bug, either. It’s one of the oldest ones there is. What’s changed is how often it’s turning up now, and how quietly it slips past teams that are moving fast enough to mistake “it works” for “it’s safe.”

A few numbers I couldn’t ignore

I don’t love leaning on statistics in something this personal, but a handful of them were hard to look past.

One assessment of vibe-coded apps from earlier this year found that more than nine in ten had at least one vulnerability that traced back to the AI either hallucinating something that didn’t exist, or simply skipping security context a human would have added without thinking twice. Georgia Tech’s security lab, which tracks vulnerabilities caused specifically by AI-generated code, saw a sharp jump in March — from single digits in January to dozens just two months later. A separate scan of thousands of publicly deployed vibe-coded apps this spring found that around four in ten were leaking sensitive data of some kind — medical records, financial details, internal company files.

None of this means vibe coding is a bad idea. It means the industry is finding out, in public, what happens when speed gets ahead of review.

So was Vembu right?

Not entirely — and I want to be honest about that instead of just declaring a winner now that the evidence tilts his way. Pichai’s argument hasn’t fallen apart either. Google is still shipping a huge share of its new code with AI help, and it hasn’t caused Google-scale disasters, mostly because Google still has thousands of senior engineers reviewing what comes out the other end. The democratization piece of his vision is real too — people who couldn’t code six months ago are shipping working products today. That much hasn’t changed.

What’s changed is where the risk actually sits. It’s not spread evenly. A well-resourced company using AI to write code, with humans still checking it, seems fine so far. It’s the solo founder, or the three-person team shipping an entire product straight from prompts with nobody looking at the plumbing, where things are falling apart. Not because the tool failed them, but because the old safety net — a second engineer looking over the first one’s shoulder — quietly vanished, and nobody put anything in its place.

Vembu wasn’t really saying AI-written code is bad code. He was saying tech debt doesn’t announce itself. It just sits there, invisible, until someone stumbles onto it — and it’s usually not the person who built it.

What I’d tell someone building something right now

If you’re a developer: none of the incidents above came from some exotic vulnerability. They came from nobody asking the boring questions — who can access this, what happens if someone sends bad input, what’s the default permission here. Those are still your questions to ask, even when the code technically isn’t “yours.”

If you’re a founder, or building something solo: speed got you to launch. It won’t get you through what comes after. A short, unglamorous review before anything touches real user data will save you a far worse week down the line.

I still don’t have all the answers here, and I don’t think anyone fully does. AI is moving fast enough that even predictions from a couple of years ago can already look off.

Maybe the real question isn’t whether AI is going to change our lives — it already has. It’s how we choose to use it, and what we owe each other while we figure that out.

I’d genuinely like to know what you think. Are we heading somewhere we’re ready for, or are we just moving faster than we can keep up with?