AI code compiles. But is it secure?

AI models that write code get better every few months. They write faster, make fewer syntax errors and solve bigger problems. But one question still lacks the answer everyone wanted: is this code secure?

On July 28, 2026, Veracode, a US application security company, released its 2026 GenAI Code Security report. The short answer: a little over half the time.

What Veracode measured

The company tracks more than 100 AI models, across four rounds of testing since the program began. The mid-2026 ranking covers 11 models. The main results:

• Generated code compiles in roughly 100% of cases.
• The average pass rate on security tests is 56%.
• Without explicitly asking for security in the prompt, the failure rate nears 44%.
• The number is “virtually unchanged since last year's report.”

In other words: the code almost always runs and is secure a little over half the time. And that half isn't improving.

A bigger model doesn't fix it

The report knocks down a few common assumptions.

• The top model in the ranking, OpenAI's GPT-5.5, passed 68% of tests. The worst, Alibaba's Qwen3.7-max, passed 50%.
• Six of the 11 models landed between 50% and 53%.
• Code-specialized models scored 51%. General-purpose models, 52%.
• Large models, with more than 100 billion parameters, scored 53%. Medium and small ones, 51%.
• Reasoning models did better: 56% versus 51%.

Language matters too. Python code passed 63% of tests. Java, only 30%, though it's the only language with consistent improvement over the past year.

Chris Wysopal, Veracode's co-founder, sums up the problem: models can be “nearly perfect at syntax” but still fail almost half the tasks.

Why this matters if you're launching

A founder building a product with AI sees the app open, the button work and the signup save. Everything compiles. It's natural to conclude it's ready.

The report shows why that conclusion is dangerous. Compiling is the easiest test. Security is invisible until the day someone exploits the flaw. And when that happens to a product holding customer data, the damage isn't just technical. It's about trust, and sometimes legal.

Many of the rescues that reach the Studio start like this: the product works, but nobody can say whether it's secure, as in the case of the Lovable showcase app that exposed 18,000 records.

Why the number won't move

The most uncomfortable data point in the report isn't the 56%. It's the stagnation. In a year, models got faster, cheaper and better at long tasks. The security rate stood still.

One reading is that models learn from code already on the internet, and a lot of that code has the same flaws. Another is that security depends on context: who the user is, what data they can see, what happens if a bad actor sends strange input. AI writes what you ask for. If you don't ask for security or describe that context, it has no way to guess.

For a founder, the practical takeaway is simple: waiting for the next model to fix it isn't a plan. The check has to be part of the process.

What we check before going live

Wysopal recommends treating AI code “like any unreviewed code: scan it, fix it and never ship it blind.” At the Studio, we use Claude Code every day. That's exactly why nothing AI writes goes to production without passing these checks:

1. Login and permissions. Who can see what. Can one user reach another's data by changing a number in the URL?
2. User input. Is every field the user fills in handled before it reaches the database or the screen?
3. Secrets. API keys, passwords and tokens stay out of the code and out of the browser.
4. Dependencies. Are the libraries the AI picked up to date and free of known flaws?
5. Prompts that ask for security. The report itself shows failures rise when you don't ask. We always ask.
6. Human review. Someone on the team reads and owns the code that goes to production. Automated analysis tools help, but they don't replace that read.

After launch, every system ships with an AI agent that runs daily, flags issues, suggests improvements and checks that the data makes sense. Checking doesn't end the day the product goes live.

What this means for you

If your product was built with AI, by you, a freelancer or another studio, ask three questions:

1. Has anyone besides the AI reviewed the code?
2. Has anyone tested what happens when a user tries to access what they shouldn't?
3. Where are the system's keys and passwords stored?

If the answer to any of them is “I don't know,” the product can work perfectly and still be exposed.

If you have a product built with AI and want to know what's underneath before you grow, our project rescue service starts with that discovery.