# Vibe coding: who pays when AI-built software breaks

> A social network built without writing a line of code that leaves 1.5 million API keys open, tens of thousands of AI assistants anyone can reach, young developers who no longer get hired, and AI that we pay for below cost today. Vibe coding is fine for a little tool at home. The trouble starts when other people use the software, and when nobody can read it any more.

- URL: https://mautoblog.com/en/posts/vibe-coding-risks-security-skills-ai-costs-2026/
- Language: en
- Published: 2026-11-15
- Author: Matteo
- Categories: Software Development, Artificial Intelligence
- Tags: AI-Assisted Development, Cybersecurity, Career, Local AI, Hardware
- Translation (it): https://mautoblog.com/posts/vibe-coding-rischi-sicurezza-competenze-costi-ia-2026/

---

At the end of January, Moltbook was the project of the moment: a social network where the ones posting, commenting and voting are AI agents, not people. Its founder, Matt Schlicht, made a point of saying so: "I didn't write a single line of code for Moltbook."

It showed. On January 31, Wiz researchers found that anyone could read and modify the whole database. Inside were about 1.5 million agent access tokens, 35,000 email addresses and more than 4,000 private conversations, some with OpenAI API keys in plain text ([Wiz](https://www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys)). The database key sat in the page's JavaScript, which is normal with Supabase, but the rules saying who can read what were missing. The team locked everything down within hours of the report.

It isn't an isolated case, and it isn't even the most serious part of this story. The point is who pays when software built without being understood breaks. Today, tomorrow and a few years from now.

---

## The toy and the product

I'll start with something I genuinely believe: vibe coding, asking AI for code and accepting it without reading it, is perfectly fine for some things. You make a little tool to rename your photos, a catalogue for your record collection, a small video game. If it breaks, the damage stays with you. Often it isn't even the AI's mistake: you asked for the wrong thing, and it did exactly what you asked. That happens to programmers too, with their colleagues.

The trouble starts when the software leaves your house. On YouTube I see more and more videos of people with no stated skills building, in an afternoon, apps that hundreds or thousands of people then use. With logins, personal data, sometimes payments. None of them can say where that data ends up or who can read it, because nobody decided: the model decided, or nobody did.

Moltbook is the most famous example, not the first. In 2025 a researcher, Matt Palmer, looked at apps published with Lovable, one of the most popular platforms for building applications by describing them in words. Of 1,645 apps checked, more than 170 let anyone read their users' data: addresses, personal debts, API keys ([Superblocks](https://www.superblocks.com/blog/lovable-vulnerabilities)). Same mistake as Moltbook, same database, same missing rules. It became an official vulnerability, CVE-2025-48757. Lovable replied that setting up those rules is up to whoever publishes the app, and has since changed its defaults. Technically it's right. But someone who publishes an app without knowing how to program doesn't even know those rules exist.

## The assistant that opens your front door

One case worries me even more, because it reaches into people's homes. [I wrote about OpenClaw in February](/en/posts/openclaw-ai-agent/): it's the AI agent that runs on your computer or on a server around the clock and can read your email, answer your messages, control your smart home. It's incredibly convenient, and that's exactly why it's dangerous.

In February SecurityScorecard counted more than 40,000 OpenClaw installations reachable from the internet, on more than 28,000 IP addresses. More than 12,000 of them could be taken over completely from a distance, and 549 were already linked to earlier attack activity ([Infosecurity Magazine](https://www.infosecurity-magazine.com/news/researchers-40000-exposed-openclaw)). In January the researcher Jamieson O'Reilly had found hundreds with their control panel wide open: API keys, Telegram and Slack tokens, months of conversations.

Think about what that means. An assistant that reads your email and turns on your lights is also a way in for anyone who finds it. And the risk isn't only yours: a compromised machine becomes a tool to attack others, your contacts, your network, the people around you. Whoever set it up by following a tutorial has no idea what "exposing a port to the internet" means. They just asked the AI to make it reachable from their phone, and the AI did.

## It happens to Amazon too

You might think this is an amateur problem. Not only. In February the Financial Times reported that in December an AWS system used by customers went down for 13 hours after some engineers let Kiro, Amazon's coding agent, make changes. Kiro reportedly decided to "delete and recreate the environment" from scratch. Amazon disputes that account: it speaks of human error, not AI error, of a single service involved and of permissions broader than they should have been. After the incident it still made peer review mandatory ([The Register](https://www.theregister.com/2026/02/20/amazon_denies_kiro_agentic_ai_behind_outage/)).

Let's take Amazon's version at face value: it was human error. That's precisely the point. The agent does what its permissions allow, and someone has to know which permissions to give it. The developer Carl Brown, quoted by Ed Zitron in a long interview on The Diary of a CEO, sums it up: AI makes the easy things even easier, and the hard things even harder. If the code of a company like Amazon needs a person checking it, imagine the app of someone who has never seen a database.

## The generation we're not training

This is the part that worries me most, and it's about my own trade.

How do you become a senior developer? By doing junior work for years. The simple pages, the boring bugs, the small changes. That's where you learn how a system really works, why that query is slow, why that check must not be removed. Today that work is done by AI, and juniors are no longer being hired.

The data is there. Erik Brynjolfsson and two Stanford colleagues follow the payroll of millions of American workers, the payroll handled by ADP. In the latest version of their study, updated in August 2026, employment of 22-to-25-year-olds in the jobs most exposed to AI is 19% lower than it would be had it kept pace with their peers in other jobs. Experienced workers show no gap at all. And the reason isn't layoffs: young people simply aren't being hired ([Stanford](https://digitaleconomy.stanford.edu/app/uploads/2025/11/CanariesintheCoalMine_Nov25.pdf)). A year earlier the gap was 15%: it's widening.

Someone will say: it's always been this way. When compilers arrived, nobody wrote assembly any more, and that's fine. But there's a huge difference. A compiler always gives the same result: you can trust it without reading what it produces. AI code can't be trusted like that. It has to be read, every time, by someone who understands it. If we stop training the people who can read it, in ten years we'll have systems written by machines and nobody able to work out why they broke. I saw it on a small scale [a few days ago, with a server of mine that wouldn't boot](/en/posts/us-flights-ground-stop-faa-software-update-2026/): when things break, you need someone who knows where to look.

## A price that isn't the real one

And here comes the bill. Today we pay little for AI. Twenty euros a month, two hundred for the professional plans. [I've already written](/en/posts/ai-economics-big-tech-ai-labs-losses-2026/) that this price doesn't cover the costs: OpenAI and Anthropic lose billions, and flat-rate plans lose money on their heaviest users. It's the same playbook as Uber and many platforms: low prices to win the market, get everyone used to it, become indispensable. Then prices go up.

Companies are already starting to see the real price. Uber itself, in April, had used up its whole 2026 budget for Claude Code in four months: use among its 5,000 engineers had gone from 32% to 84% in three months. In May its chief operating officer, Andrew Macdonald, admitted it is hard to link all those tokens to features that are more useful for customers ([Fortune](https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/)).

Now put the pieces together. Companies lay people off or stop hiring, because there's AI anyway. The code is written by AI and nobody in-house can really read it any more. At that point there's no going back: you pay what they ask. And if the AI bubble really bursts (Zitron says it could happen as early as 2027, although many of his past predictions [haven't come true](https://danluu.com/zitron/)), some of these companies could disappear. The ones left will be able to charge whatever they like.

## What about models at home?

The obvious answer is: then use open models, locally. I've written about it often, from [DeepSeek with Ollama](/en/posts/install-deepseek-locally-with-ollama-guide-2025/) onwards, and I still think they're the real alternative. They put a ceiling on what the big companies will be able to charge.

But there's a catch. A local model that comes close to the big labs' models needs 64, preferably 128 GB of memory. The PC, tablet or phone people have at home has 8 or 16. And memory costs a lot more today, for a precise reason: Samsung, SK hynix and Micron have moved production towards memory for AI data centers. According to TrendForce, DRAM prices rose 55-60% in a single quarter as early as late 2025, and IDC expects the shortage to last into 2027 ([The Register](https://www.theregister.com/2026/01/06/memory_firm_profits_up_as/)).

I can see it with my 12.9-inch M1 iPad Pro. I bought it on Black Friday 2022, on offer with the pencil, for 590 euros. Today on Back Market the same model, refurbished, starts at 629 euros ([Back Market](https://www.backmarket.it/it-it/l/ipad-pro-12-9/0522ce5c-86fa-45e4-86fc-89018d19f2d6)): more than I paid for mine, new and with the pencil. A four-year-old tablet worth more than when I bought it. With electronics, it usually goes the other way.

So the circle closes in a rather mocking way. The alternative to the cloud exists, but it needs hardware that keeps getting more expensive, because the cloud itself is buying up all the memory. And it needs skills to install it and keep it secure: exactly the skills we're no longer training.

## What I think

I'm not against using AI to program. I use it every day, and this blog has a whole diary [about how I work with agents](/en/posts/contextdoc-diary-vscode-extension-conceptual-tests-2026/). But I use it knowing how to read what it writes, and deciding myself what goes in and what doesn't.

What I'd ask is simple:

- **If it's for you, do what you like.** The little tool at home, the game, the script: vibe coding is a great way to start, and even to learn, if you then read what it wrote.
- **If other people use it, you need someone who understands it.** Personal data, internet access, payments: either you learn, or you get it checked by someone who knows. It isn't bureaucracy, it's respect for the people who trust your app.
- **An assistant that controls your home shouldn't be exposed to the internet** unless you know exactly what you're doing. Better a private VPN like Tailscale, [as I explain in the OpenClaw guide](/en/posts/openclaw-complete-guide-channels-security-part-2/), or nothing.
- **Companies should keep hiring juniors.** Not because they're kind, but because in ten years they'll need seniors, and you can't buy those ready-made.

AI is cheap today because someone is picking up the bill for us. When they stop, only one thing will matter: still having people who know what's inside.

