AI writes code four times faster. And bugs ten times faster.

The productivity claims about AI coding assistants are real and measurable. So, it turns out, is the other half of the ledger, and 2026 was the year enough code shipped for somebody to count it. None of this is an argument against the tools. It is an argument about who checks the output, and what happens when nobody does.

What the 2026 studies measured

The Veracode figures are from controlled security-sensitive coding tasks, not from production repositories, so they measure what models produce when asked rather than what ships after review. The CVE counts are confirmed disclosures only; the researchers estimate the real figure is five to ten times higher across open source.

The two numbers that belong together

Across Fortune 50 engineering teams, developers using AI assistance produced commits at three to four times the rate of their peers and introduced security findings at ten times the rate. Both halves are real. Put together they describe not a faster team but a team accumulating a particular kind of debt faster than it can be reviewed.

Nothing here says the code does not work. It says it works and is exploitable, which is a much harder failure to notice, because the thing you test for is whether the feature does its job. Cross-site scripting and log injection do not show up when the form submits correctly.

Why the models fail in this specific way

A model produces the most probable code for the request, and the most probable code on the internet is the tutorial version — the one that demonstrates a concept without the escaping, validation and parameterisation that production needs. Ask for a search box and you get a working search box, written the way search boxes are written in examples.

That is also why the failure rate is so uneven by language: Java came out worst at 72%, tracking the age and volume of the example code available for it. The model is not making a mistake by its own standards. It is reproducing the distribution it learned, and that distribution is mostly demonstrations.

What this means if you bought a cheap build

If a very low quote came with a very short timeline, it is worth knowing whether anyone reviewed the output for anything other than "does it look right". The visible layer — pages, forms, layout — is exactly the part AI does well, so a site can look entirely finished and carry a form that will accept a script tag.

The cheapest moment to find this is before launch; the most expensive is after a breach discloses it for you. A technical audit covers it directly, and what an AI-generated website actually gets you is the wider version of the same question.

What responsible use looks like

The teams getting the productivity without the debt are doing three unglamorous things: automated security scanning in the pipeline so findings surface at commit rather than at audit, human review of anything touching input handling, authentication or data access, and an explicit rule that generated code is a draft.

That is not a brake on AI, it is the condition under which the speed is worth having. We use these tools daily and hold the same rule internally — the reason we build with Next.js is partly that its defaults escape output by design, which removes an entire class of this problem before anyone types anything. The other half of the same year is that attackers adopted the same tooling, and faster — black hat hackers now run agent swarms is that story.

Sources

Checked September 2026. Controlled studies and disclosure counts — see the table footnote for what each does and does not measure.

• Veracode — Spring 2026 GenAI code security update ↗

  Over 100 large language models on security-sensitive tasks: 45% of samples introduced an OWASP Top 10 flaw, 86% failed XSS defence, 88% log injection, Java worst at 72%.

• Cloud Security Alliance — Vibe coding’s security debt: the AI-generated CVE surge ↗

  The CVE progression of 6 → 15 → 35 per month across January to March 2026, and the estimate that confirmed disclosures understate the real figure five- to tenfold.

• Infosecurity Magazine — Researchers sound the alarm on vulnerabilities in AI-generated code ↗

  Coverage of the commit-rate versus finding-rate comparison across enterprise engineering teams quoted in the first section.