What AI can copy, and what it can't

What you hand out can be studied

Once a program leaves your hands, whoever receives it can open it, see how it works and look for flaws. That has always been true. What's new is the cost.

In October 2026, a researcher published what he calls the first jailbreak for the Nintendo Switch 2 and said he built it in about two weeks with help from an AI coding tool. The outlets we read had not verified it independently.¹ That same month, PHP fixed two critical flaws, and a security firm found one of them using an AI model.²

This is not only about game consoles or programming languages. Your business app, a file your customers download or the code that runs in your visitors' browsers are in the same position: they can be studied.

What can't be copied that way

Some things nobody can take apart because they never get to hold them:

• What stays on your server. If the business rule, the calculation or the key lives on your server and the customer only sees the result, there is nothing to open.
• Information encrypted with sound cryptography. Without the key, encrypted data can't be read, however fast someone tries.
• Keys stored in dedicated hardware. The large cloud providers offer hardware modules built to hold keys.³

OWASP's mobile app security standard, from a nonprofit foundation that is a reference in software security, puts it this way: security must rely on verifiable design, strong cryptography and server-side validation.⁴

API keys, the loot of the moment

An API key is the password a program uses to get into a service, for example your company's AI account. Whoever holds it uses the service in your name and on your bill.

On October 20, 2026, the security firm Flare published an analysis of more than 100,000 Telegram channels and 100 dark web forums. Mentions of open AI models with their safeguards removed went from 11 to 93 per 30-day window between January and October 2026. According to Flare, most of what is sold looks like resold or stolen access, that is, other people's API keys. Flare also documents seven AI-assisted attacks between May and October 2026; one of them reached at least nine South Korean banks. The study itself warns that its counts are a floor and a direction, not a census.⁵

Five practices for your business

1. Secrets stay on the server. Never inside an app, a file you hand out or your site's code. OWASP notes that many organizations have secrets written into their code in plain text.³
2. Keys that are worth little if they leak. Each key with the minimum permissions it needs, an expiration date and regular rotation, so a stolen key only works for a short time.³
3. Watch the usage on your AI accounts. Spending you don't recognize can be the first sign of a key in someone else's hands.
4. Obfuscation helps, but it doesn't replace the above. Making code hard to read raises the cost of studying it, but OWASP warns it must not be used as a substitute for a sound security architecture.⁴
5. Legal protection counts too. Contracts, licenses and trademark registration protect what technology can't reach. Set that up with your lawyer.

If a key has already leaked

Revoke it and create a new one, review that account's recent usage and change related passwords too. OWASP recommends revoking credentials that are no longer needed.³

In our digital security service we review where your business keys and secrets live, with your written authorization and without asking for your passwords.

Ask us how we do it for your business, on a free 30-minute call.

General information, not technical or legal advice.