Instructing AI: alarm system or key under the doormat?

Today, you only need to describe your idea to an artificial intelligence to get a website or app within a few hours. It's called vibe coding. The result often looks great. But a good-looking site isn't necessarily a safe one.

Picture two identical houses. The first is protected by an alarm system linked to a monitoring centre. In the second, the key is under the doormat. From the street, you can't tell the difference. To a burglar, it's obvious.

The instructions you give the AI are your alarm system. A request without instructions is the key under the doormat.

The key under the doormat: what happens without instructions

When you simply ask "build me a booking app", the AI has one goal: making it work. It doesn't ask who will be able to see your customers' information, or what the law says. The result: many apps look like they work perfectly, but leave their customers' personal data within anyone's reach: names, email addresses, sometimes bank or medical details. No genius hacker is needed to grab it: the door is already open.

Then there's the law. If you publish the app, you are responsible for your customers' data, not the AI. And a nice design doesn't replace a privacy policy, a cookie banner or legal notices.

The alarm system: giving the AI instructions

An alarm is installed before the break-in, not after. With AI it's the same: the instructions are given from the very first request, and stay written down for the rest of the project. They come down to four points:

• Context: who will use the app, what information it will store, and in which country. An AI that knows it's handling health data in Belgium doesn't work the way it would for a simple brochure page.
• Security rules: no passwords hidden in the page, each customer sees only their own information, nothing is saved without being checked.
• Legal obligations: privacy policy, cookies, legal notices, accessibility.
• The final test: check with two different accounts that one customer can never see another's information.

An alarm isn't enough on its own

A word of caution: giving the AI instructions greatly reduces the risk, but doesn't remove it. The AI can forget a rule, and every new change can reopen a door. Like an alarm, it needs to be checked and maintained: a professional reviewing the work, tests, and follow-up after launch.

Why the bill arrives in the years ahead

A poorly protected app doesn't disappear: it stays online, and it ages. Hackers don't even need to target you personally: bots sweep the internet around the clock looking for open doors. And stolen data can never be recovered. You can change a password, but not a date of birth, an address or a medical record. This information is resold and used for years in scams and identity theft.

The more apps get published in a rush, the more open doors there are, and the more cyberattacks succeed. The problem isn't AI: it's the lack of instructions.

AI builds the house in a weekend. It's up to you whether it installs the alarm or slips the key under the doormat.

Sources

• Veracode — 2025 GenAI Code Security Report
• Escape — 2k+ vulnerabilities in apps built with vibe coding platforms
• Techzine — Moltbook database exposes 35,000 emails and 1.5 million API keys
• Bruni et al. — Benchmarking Prompt Engineering Techniques for Secure Code Generation (FORGE 2025)
• Security Degradation in Iterative AI Code Generation (IEEE-ISTAS 2025)
• RGPD — Règlement (UE) 2016/679, articles 32, 33 et 83