AI Writes Half Your Code Now. 44% of It Ships With a Vulnerability

If your company has adopted AI coding tools, AI now authors roughly half of all the code your teams commit to production, on average. And 44% of AI code-generation tasks introduce a real security vulnerability, the kind a human reviewer would flag as risk (Veracode, 2026 GenAI Code Security Report). This isn’t a warning about the future: CVEs (publicly catalogued vulnerabilities) attributed to AI-generated code went from 6 in January 2026 to 35 in March, nearly a sixfold jump in two months (Georgia Tech, Vibe Security Radar project). The problem isn’t that companies are slow to adopt AI for coding. It’s that they’re adopting it far faster than they’re auditing it.

From “Writing Faster” to “Writing Half of Everything”

The average security pass rate for AI-generated code sits at 56%, barely moved from 55% in the previous report — unchanged even though the underlying models have changed completely in a year (Veracode, 2026). Among the models tested, GPT-5.5 leads at a 68% pass rate, but more than half of all models tested land between 50% and 53%. Reasoning models do somewhat better than non-reasoning ones (56% vs. 51%), and models marketed as “coding-specialized” aren’t meaningfully better than general-purpose ones (51% vs. 52%). In other words: no single model solves this yet, no matter how much you pay for it.

The 44% Isn’t Evenly Spread. That’s the Part No One Mentions

What almost no one says out loud is that this 44% isn’t a uniform risk. AI handles the vulnerability classes it has effectively memorized well — SQL injection, 83% pass rate; cryptographic algorithms, 87% — but fails systematically at the two classes that almost never get caught in manual review because they don’t break anything visible: cross-site scripting (XSS), at just 15% pass rate, and log injection, at 12% (Veracode, 2026). The real risk isn’t in the issue everyone already talks about (SQL injection, which every team already reviews); it’s in the one nobody looks for because it never throws an error on screen.

The CVE Curve Nobody Is Watching

Georgia Tech’s Systems Software & Security Lab launched the Vibe Security Radar project in May 2025 to track vulnerabilities originating from roughly 50 AI-assisted coding tools (Claude Code, GitHub Copilot, Cursor, Devin, and others). The result through March 2026: 74 CVEs confirmed as directly attributable to AI-generated code, climbing 6 (January) → 15 (February) → 35 (March). The researchers themselves estimate the real figure is 5 to 10 times higher across the broader open-source ecosystem, putting the actual problem somewhere between 400 and 700 exploitable flaws (Infosecurity Magazine, reporting on the Georgia Tech research).

The case they point to is the OpenClaw project: over 300 published security advisories, and researchers could only confirm around 20 cases with a clear AI signature, because the metadata that would reveal the code’s origin is routinely stripped before the commit lands. As project founder Hanqing Zhao put it: “If that commit has an AI tool’s signature on it, we flag it. But teams that do code review aren’t going to catch everything when half the codebase is machine-generated.”

Why This Is No Longer Just a Developer Problem

The context makes it worse: software vulnerabilities are already the number one breach entry point, at 31% of cases (Verizon, 2026 Data Breach Investigations Report, cited in Veracode’s report), and high-risk vulnerabilities are up 36% year over year (Veracode, 2026 State of Software Security). That’s not a coincidence alongside the rise of AI-generated code — it’s the same curve.

This links directly to two risks we’ve already covered separately: non-human identities (an AI coding agent is, in practice, one more account with access to your repository and sometimes to production) and least-privilege permissions for AI agents: if your coding copilot can deploy directly without review, the problem isn’t just the vulnerability it writes, it’s the access it has to ship it. The same applies if that agent connects to your internal systems through MCP: the entry point isn’t just the code, it’s the connection.

What to Do This Week

• Measure what share of your code is already AI-written, not by guessing: most companies don’t have this number, and it’s the first question to answer before deciding how much extra review is needed.
• Review XSS and log injection first, not just SQL injection: these are the two points where AI fails most (15% and 12% pass rates) and the ones least likely to be caught on a quick read.
• Treat every AI coding tool as an identity with access, not just a text editor: what can it deploy without anyone reviewing it?
• Don’t assume a “better” or “coding-specialized” model fixes this: no model tested exceeds a 68% pass rate, and specialized models barely outperform general-purpose ones.
• Ask your team how many security advisories your own code is already carrying, before an external auditor — or an attacker — does it for you.

How We Help

At Optima Quantum, we start with a cybersecurity audit that measures, with data instead of guesswork, how much of your code and deployments already depend on AI tools with no extra security review, and what real access those tools have to your systems. From there, we design the review and governance process you’re missing with you, as part of our cybersecurity services — without selling you “more AI” as the automatic fix for a problem AI itself is creating. If your team already codes with AI and you don’t know how much of that code has actually been reviewed, contact us and we’ll look at it together.

Frequently Asked Questions

Does 44% of AI-generated code really ship with a vulnerability? According to Veracode’s 2026 report, 44% of the AI code-generation tasks tested introduced at least one real security vulnerability, and the overall average pass rate is 56%.

Which vulnerability types are the most common? The ones least likely to be caught by hand: cross-site scripting (just 15% pass rate) and log injection (12%). SQL injection and cryptographic algorithms, by contrast, are handled well (83% and 87%).

Are there already real catalogued vulnerabilities (CVEs) because of this? Yes: Georgia Tech’s Vibe Security Radar project confirmed 74 CVEs directly attributable to AI-generated code between January and March 2026, climbing from 6 to 35 in that period.

Does using a pricier or “coding-specialized” AI model solve the problem? Not decisively: the best model tested (GPT-5.5) only reaches a 68% pass rate, and coding-specialized models don’t meaningfully outperform general-purpose ones.

Does this only affect companies that have already had an incident? No: 31% of breaches in 2026 already start with a software vulnerability, and high-risk vulnerabilities are up 36% year over year. The risk exists even before any visible incident happens.