Is AI code secure? Not by default. AI-generated code usually works, but studies show a large share of it contains security flaws that nobody asked for and nobody notices until later. Used with the right checks, AI is a powerful assistant. Used blindly, it can quietly add vulnerabilities to your product.
If you build websites or apps, you have probably used an AI coding tool, or your team has. At SparkSupport, we work with AI-assisted development every day, and this is the question we hear most. So we reviewed the most-discussed research and expert opinions on AI code security and put the answers in one place, in plain language. You should not need ten tabs open to understand this.
Is AI-generated code safe? The short answer
AI code works, but “works” is not the same as “secure”.
AI can write a login page, an API or a checkout flow in minutes. It is very good at making the feature run. It is much weaker at remembering the protections nobody mentioned in the prompt, such as permission checks, input validation and safe handling of secrets. Without a human or a process to catch those gaps, they ship.
The good news is that this is fixable. The rest of this guide explains what the research says, why AI code goes wrong, what well-known experts think, and how to use AI safely in your web and app projects.
Why everyone is asking about AI code security
AI coding is no longer a side experiment. Google’s 2025 DORA report found that about 90% of software professionals surveyed use AI at work. The same report found that around 30% of them have little or no trust in the code it produces.
That gap, “we all use it, but many of us don’t fully trust it”, is exactly why AI code security matters now.
What the research says about AI-generated code security
Several studies have tested AI-written code. The numbers differ because each team measured differently, but the message is consistent.
- Early warning (2021): Researchers at New York University tested GitHub Copilot with 89 security-related scenarios. Roughly 40% of the generated programs were vulnerable.
- A large study across 9 AI models: Published in the journal Empirical Software Engineering, this study generated about 331,000 programs with nine different models and checked them using formal verification, a mathematical way of proving where code can fail. At least 62% of the programs had vulnerabilities, and the models made very similar mistakes. The authors concluded that AI output needs proper risk assessment and validation before production use.
- Design flaws and known vulnerabilities: A study cited by the Cloud Security Alliance found that 62% of AI-generated solutions contained design flaws or known security problems, even with the latest models.
- Real-world projects: Researchers who studied Copilot-generated code that developers actually used in GitHub projects found security weaknesses in about 29.6% of snippets, across 38 different weakness types.
- “Works” is not “secure”: A 2026 benchmark called SusVibes tested AI agents on 186 real feature requests. The best setup got 57% functionally correct, but only 11.8% were secure. Among the solutions that worked, 79.3% still contained a vulnerability.
- A fair counterpoint: A large analysis of public GitHub repositories found that about 88% of AI-generated code had no identifiable, catalogued vulnerability. So the picture is not “everything is broken”. It is “a meaningful share is risky, and you will not know which share unless you check.”
An honest note: several of these studies used the C language or lab setups, while web apps are usually built in JavaScript, Python, PHP or similar. The exact percentages will not match your project. What carries over is the pattern: AI repeats common, well-known mistakes, and they hide behind code that looks fine.
Why AI-generated code ends up insecure
There is no single bug. It comes down to a few simple reasons.
1. It learned from the internet, mistakes included.
AI models learn from huge amounts of public code. If unsafe habits appear thousands of times online, such as building a database query by gluing user input into a string, the AI repeats them.
2. It takes the shortest path to “working”.
If you ask for something vague, the AI picks the quickest solution, even a dangerous one. For example, it may use a shortcut that runs whatever text a user types, which can let an attacker run their own commands.
3. It forgets things nobody asked for.
Security is often about what is missing: a permission check, input validation, a limit on requests. If your prompt says “build a checkout endpoint” and says nothing about checking who is calling it, the AI will often skip the check. The feature still works, so nobody notices.
4. It makes quiet logic mistakes.
A small rewrite can change the meaning of a rule. Imagine a check that works for users with one role but fails for users with two, accidentally giving them too much access. The code looks correct and passes simple tests.
5. Knowing is not the same as doing.
Researchers at the University at Buffalo describe a “knowledge-actuation gap”: AI models often know the security rule but fail to apply it in the exact place in the code where it matters.
6. The person prompting may not spot the problem.
If you can see that the feature works but cannot judge whether it is safe, nobody is really reviewing the security. This is the biggest risk when non-specialists build apps entirely through prompts.
The most common vulnerabilities in AI code
These are not exotic. They are the oldest, best-known weaknesses in software, and AI can produce them faster than people can review them.
- Injection attacks: user input is placed directly into database queries or system commands, so an attacker can run their own.
- Broken access control: the app shows data without checking whether the person is allowed to see it. OWASP, the industry’s main web-security list, ranks this as the number one risk for 2025 and reports that every application it tested had some form of it.
- Hardcoded secrets: API keys and passwords written straight into the code, where they can leak through the repository.
- Missing input validation: the app trusts whatever users send.
- Unsafe handling of saved data: loading stored objects in a way that can let attackers run code.
- Business-logic flaws: a negative quantity in a cart, a coupon that stacks forever, a refund without an ownership check. These are the hardest to catch, because the code is technically perfect but the rule is wrong.
What tech leaders say about AI code and vibe coding
This debate is not just about studies. Some of the best-known people in software have weighed in, and their views show where the line sits between smart use and risky use.
Andrej Karpathy: the person who named “vibe coding”
Andrej Karpathy, a co-founder of OpenAI and former AI leader at Tesla, coined the term in February 2025. He described a style of coding where you “forget that the code even exists”: you describe what you want, accept what the AI gives you and paste in errors until it works.
What it means for you: Karpathy himself framed this as suited to throwaway weekend projects, not serious software. It is brilliant for a prototype. It is not a plan for an app that handles customer data or payments.
Simon Willison: where AI use becomes real engineering
Simon Willison, a well-known developer and writer on AI tools, draws a clear line. If an AI wrote every line but you reviewed, tested and understood it all, that is “using an LLM as a typing assistant”, not vibe coding. He has also called taking vibe-coded output straight into a production codebase clearly risky.
What it means for you: The issue is not whether AI wrote the code. The issue is whether a knowledgeable human understood and checked it before it shipped.
Linus Torvalds: fine for learning, risky for products
Linus Torvalds, the creator of Linux, said in a late-2025 interview that he is fairly positive about vibe coding as a way for newcomers to get started and make computers do things. But he warned it can be “a horrible, horrible idea from a maintenance standpoint” if you are building a real product.
What it means for you: Getting something running is only the first step. Keeping it secure, updated and fixable for years is the harder job, and AI-generated code that nobody understands makes that job harder.
Bruce Schneier: security is never a one-time purchase
The security expert Bruce Schneier is famous for the line “security is a process, not a product”. It is an old idea, and it fits AI coding well: no tool, however smart, makes you secure on its own. Security comes from habits, reviews and checks that keep running.
What it means for you: Do not look for an AI tool that makes you safe. Build a process around the tools you use.
The common thread
None of these voices say “never use AI”. They all say, in different words: use it, but keep a human who understands the code in the loop.
Can AI check its own code for security problems?
It helps a little, but it is not enough. The AI that wrote the code shares its own blind spots. It does not know your business rules, who should see what, or how your system fits together. Research on the SusVibes benchmark found that adding security hints to the prompt did not fix the problem.
A trustworthy check needs an outside viewpoint: security scanners that follow how data really flows, your own written rules, and a human reviewer.
How to use AI code safely in web and app development
Here is a practical checklist you can use today, whether you are building a website, a web development project or a mobile app.
- Treat AI code as untrusted code. Review it as you would code from a new, unknown contractor.
- Be specific in your prompts. Say what you need: “validate all inputs, check the user’s permissions, use parameterized queries, never hardcode secrets.” The prompt is now part of your design.
- Never paste real secrets into prompts or code. Use environment variables or a secrets manager.
- Add security checks early. Do not wait for the final review. Run static analysis, dependency scanning and secret scanning on every pull request.
- Keep humans reviewing what matters. Login, payments, permissions and anything touching personal data deserve a careful human review.
- Test the unhappy paths. Try wrong users, negative numbers, empty fields and unexpected requests, not just the happy path.
- Watch your dependencies. AI may suggest libraries that are outdated or unnecessary. Check them.
- Write down your rules. Keep your security and business rules documented so humans and tools can check against them.
- Keep humans accountable. Someone on your team should be able to explain what every important piece of code does.
- Use AI for speed, humans for judgment. Let AI draft, then let experienced people decide.
The same rules apply on mobile. If you are planning Android app development, review AI-generated code for secure storage, API calls and permissions just as carefully as you would on the web.
Build with AI, without the security guesswork
At SparkSupport, we use AI to build faster, and we also see the other side: the reviews, security checks, testing and DevOps practices that make fast code safe to run. Our view matches what the research says: the speed of AI and the judgment of experienced engineers work best together.
If you want to add AI to your product or your development process, our AI services team can help you do it with the right guardrails from day one.
Frequently asked questions about AI code security
Is AI-generated code safe to use in production?
Only after it has been reviewed and tested like any other code. Studies show that a significant share of AI code contains weaknesses that do not break the feature, so they slip past simple testing.
Is vibe coding okay for my business app?
It is great for prototypes, demos and experiments. For anything that handles customer data, payments or logins, add proper review, testing and security scanning first.
Which AI model writes the most secure code?
The large comparison study found only minor differences between models, and all of them made similar mistakes. The process around the model matters more than the model itself.
Will AI code get safer over time?
Models keep improving, but the structural reasons, such as learning from imperfect code and not knowing your business rules, do not disappear. Plan for checks either way.
Should I stop using AI for coding?
No. Use it with guardrails. Every expert quoted above supports using AI, with a human who understands and checks the result.
The bottom line
AI code is not unsafe by nature, and it is not safe by default. It is as secure as the process around it. Write clear prompts, scan early, review what matters and keep experienced people accountable. Do that, and you get the speed of AI without paying for it later in security incidents.
Building something new, or worried about an AI-built app that is already live? Contact our team and we will help you review it and set up a secure development workflow.