November 9, 2026 · 8 min read · ai, security
The Golden Age of Cybersecurity Risks
AI makes it easy to ship the software of your dreams. Most people doing it have no idea how the code works, or the holes they're leaving behind.
I use AI coding tools fairly often, though not to build full applications or to do schoolwork I’d rather skip. The simple fact is that a model can write boilerplate code much faster than I can. I can give it an outline and have method headers and comments ready in a snap. I let it check the shell scripts I write and catch spelling mistakes in my README files. It can even be a good second set of eyes when the rubber duck isn’t cutting it anymore and I’m stuck on a bug. (If that one’s new to you, Google “rubber duck debugging.”) So, to be clear, I’m not telling you to stop using AI tools. I want to warn you about a pattern I keep seeing online, and even with friends.
The temptation to get your dream software from a single prompt is strong, and plenty of people are giving in to it. The problem is architecture. Most people who are “vibe coding” have no foundation in software engineering, cybersecurity, or DevOps, and that’s where the danger is. When someone with no software background starts building an app in an AI tool and it tells them to go here, get this API key, paste it into this hosting service, they follow the instructions. Unfortunately, they’re blithely unaware that what they’re doing could put their own data at risk, along with the data of anyone they share the app with, and even open holes in their own devices.
Working and safe are different tests
Once everything looks right, people ship the app and start using it. The trouble is that all they’ve tested is what we call the happy path. The button you wanted is there. You click that button and the form submits, the page loads, the data shows up, whatever you wanted it to do it now does. AI models have gotten extremely good at passing these surface-level tests. But what happens when a service returns a 500-class error, or a request times out and breaks the chain behind it?
When you try to “one-prompt” an app, the AI makes a single large step to fulfill the outlined request in your prompt. This includes the unverified plan it makes, then it writes the code it wants, and shows you the result. From there you’re only tweaking. That causes a whole slew of problems, because AI models are not good systems architects—the person who decides which services to use and how they fit together. Apple’s researchers showed in The Illusion of Thinking that “reasoning” models hit a wall as problems get more complex. Past a certain point their accuracy collapses entirely, even when they have “thinking” time to spare. And as a conversation grows, models get worse at using information buried in the middle of it, as the Lost in the Middle study found. A long reply that took a while to generate is not inherently a better one.
Security is where this shows. Veracode ran more than 100 models through 80 coding tasks that each had a secure and an insecure way to solve them. In its 2025 report, the models picked the insecure approach in 45% of cases. Its 2026 follow-up found the average security pass rate stuck at 56%, nearly unchanged from the year before, even as the code itself got more reliable. Models built specifically for coding did no better than general-purpose ones. So yes, the code runs… but there are big gaping holes in the application.
What do these holes actually look like?
These are some of the most common mistakes in web security, and AI repeats them because it learned from public code that’s full of them.
- Secret keys in the wrong place. API keys and database passwords pasted into front-end code or pushed to a public GitHub repo. Anyone who opens the browser’s developer tools or finds the repo can read these keys. Now they have access to your personal information through the service or if it’s say an AI API key then they can run up your token bill at no charge to themselves because it will come back to your account.
- Databases with no access rules. Hosted databases like Firebase and Supabase let you control who can read and write each record. If nobody sets those rules, any logged-in user, or sometimes anyone at all, can read everyone else’s data. Well, if you’re taking people’s information that is private and turn around and hand everyone in the world an open copy, you may find yourself in a little bit of legal trouble.
- Missing permission checks. The app checks that you’re logged in but not that the record you asked for is yours. Change one number in the URL and you’re looking at someone else’s account. You’re logged in, so why not just see everyone’s account information, right?
- Unchecked input. Whatever a user types goes straight into a database query or onto the page, which opens the door to SQL injection and cross-site scripting. People can maliciously target these services and without proper safeguards they can easily get all your stored data.
- Packages nobody vetted. The AI suggests a library that’s outdated, or one that doesn’t exist at all. Attackers have started publishing malicious packages under the names AI tends to make up as well, potentially leading to even higher rates of cyber attacks. The other problem here is that the older versions often have known vulnerabilities that newer versions have fixed.
Why beginners can’t see it
Unfortunately for the vibe coder, there is not a big blazing red sign that says DO NOT DO THIS when they try to push these error-filled apps. There’s no error message, and there is no broken button, but that doesn’t mean everything is fine. The app works perfectly for you because you’re using it the way it was meant to be used. The person who finds the hole is someone using it the way it wasn’t meant to be used, and they may not tell you about it! The hackers planting seemingly safe packages certainly don’t care that you have a vulnerability, they want money!
It’s impossible to review code that you don’t understand, and if you can’t read what the AI wrote, how could you tell a secure login from one that just trusts whatever the browser tells it? I have seen people ask AI to “check for security vulnerabilites” or to “be a DevOps engineer and look through this code” but this is not a solution. At most, we’re putting bandaids on bullet holes. AI is tailored to agree with the user, which has its own ethical implications in my opinion, but it will tend to just say that you are correct. So, while “is this secure” may help a little, it certainly will not solve the issues. The people that will end up paying for this in the long run are you as the developer, especially legally, but also the user because once the data is out there, it’s there for good.
If you’re building without a dev background
I don’t want you to feel limited by a lack of coding knowledge, so here are a few ways that you can improve (and that doesn’t mean guarantee) security and data integrity in your app:
- Keep the secrets to yourself. Don’t let anything that starts with
sk_,key, orsecretget into your front end or a public repository. If your keys are out there in a public repo, the best thing to do is to make the repo private while you fix it. You will also need to revoke the key and issue a new one with the provider, as well as scrub the history of the repo, not just current files. - Use pre-built login services. There is no need to reinvent the wheel. We have lots of secure login systems, just use a secure authentication service, don’t let AI one shot code an auth system. Firebase, Supabase, Clerk, and Sign in with Apple or Google have already solved this. If you’re using AWS, you can use Cognito for secure logins.
- Use access rules before launch. The idea here is that if user A logs into the app, then they can’t see user B’s data. They would have to login as user B to see that data.
- The principle of least privilege. Don’t give every user that signs up admin or editor permissions, just give them the bare minimum they need to do their job (e.g., read-only access).
- Scan your dependencies. You can run tools like
npm audit, or the Dependabot alerts on GitHub to find outdated or vulnerable packages, then you can update them. - Ask AI to do things in rounds. Don’t let the AI go from one single prompt to a whole app. A better way to build is in steps. Let it set up the project directory, then let it go through and write the method headers, then it can fill in the basic functions, so on and so forth. While it may take longer overall, the end result will be much better. This method allows for human review between steps too to catch early architecture bugs.
- Ask for a specific review. Don’t favor vague prompts like “is this secure?” Instead, try directed prompts that give one task (or if you’re really in the weeds, ask the AI to create a checklist of the entire codebase sweep and to go through each bullet one at a time auditing the code for each individual thing it lists). The models perform better with a narrower scope and a larger number of cycles in my experience.
- Get a human to look before real data goes in. If the app will hold personal information or take payments, pay someone who does DevOps or cybersecurity for a living to review it first. It costs far less than a breach will cost you in the long run.
If you’re a developer
Treat AI output like a pull request from someone you’ve never worked with. Read every line before it merges, run the same scans you’d run on anything else, and don’t let speed become the reason review gets skipped.
I’ve been building and running systems for clients since 2018, and most of the work that matters is the part nobody sees. Without solid backend systems, failures are bound to happen. In my experience, a client will always find some random, niche, never seen before path that breaks your code in a way you never thought of and definitely didn’t test for. AI is very good at the part people do see, though, so let it make that pretty front end for you (but I do still prefer human web design and if you’re looking, Digital Dolce is the place to go!)
The bottom line is this: We want to share our dream software, but we cannot allow the bugs and security holes to destroy our personal data along the way. Hold on tight to your software engineers, DevOps managers, and your testers too!