AI-Generated Code Security Stuck at 56% as Vibe Coding CVEs Rise

A flat line in a steep market

AI coding tools keep getting more capable, but the code they write is not getting safer. Veracode's 2026 GenAI Code Security Report puts the average security pass rate across the models it tested at 56%. That is barely above the 55% recorded in its first report 4. Put the other way, roughly 44% of code generation tasks produced output containing a known vulnerability 42.

The test is narrow on purpose. It asks whether a model's code passes a standard security check when the developer gives no explicit instruction to write secure code 2. Veracode has run four testing snapshots and tracked more than 100 large language models 2. Its central finding is that "syntax is solved, security is not." Models now reliably produce code that compiles and runs, but security performance has not followed the same upward curve 4. The company also says that greater model capability does not automatically bring better security 4.

Some figures differ slightly in the reporting. A Cloud Security Alliance research note describes 45% of samples introducing OWASP Top 10 vulnerabilities, against Veracode's own 44%. The CSA note also says the pass rate has not improved across testing cycles from 2025 into early 2026, "despite vendor claims to the contrary" 1. Tech Insider's coverage gives two publication dates for the report, July 28 and August 1 2. Neither difference changes the conclusion. Across every snapshot, the needle has barely moved.

From lab benchmarks to real CVEs

Benchmarks are one thing. Disclosed vulnerabilities in shipped software are another, and that number is climbing. Georgia Tech's Vibe Security Radar project counted at least 35 CVEs disclosed in March 2026 that it attributed directly to AI-generated code. The count was six in January and 15 in February 3.

The team monitors about 50 AI-assisted coding tools, including:

• Claude Code
• GitHub Copilot
• Cursor
• Devin
• Windsurf
• Aider
• Amazon Q
• Google Jules

In total, the project has confirmed 74 CVEs tied to these tools 3. Claude Code appeared most often. Researcher Zhao cautioned that this mostly reflects how the tool "always leaves a signature," which makes its output easier to attribute than that of competitors 3.

The attribution problem cuts both ways. The dashboard captures only cases that leave metadata traces, so the true number is "almost certainly higher," particularly in open-source projects 3. The Cloud Security Alliance cites researcher estimates that the real count could be five to ten times greater across the wider open-source ecosystem 1. Treat that range as an informed estimate, not a measurement. Still, a rise from six to 35 attributable CVEs in three months is hard to explain away.

Velocity is the multiplier

The most consequential figure may be about volume rather than per-line quality. According to research on Fortune 50 enterprises cited by the CSA, AI-assisted developers commit code three to four times faster than their peers. They also introduce security findings at ten times the rate 1. The CSA calls the result a security debt that grows faster than organizations can pay it down 1.

Veracode makes a similar point. Security performance has stayed flat while the amount of AI-generated code entering pipelines has surged 4. A constant 44% flaw rate applied to a much larger body of code means far more total exposure, even though the rate itself has not worsened.

The tooling keeps advancing anyway

None of this has slowed the release cycle. OpenAI launched GPT-5.5 on April 23, 2026, promoting stronger agentic coding, and made it available in Codex on supported ChatGPT plans. Anthropic is positioning Claude Opus 4.7 for hard, long-running coding work 5. Vendors sell these tools mainly on capability and autonomy. Veracode's data suggests those gains do not carry over to security on their own 4.

What to take from it

The evidence points to a structural gap, not a temporary lag. The benchmark pass rate has held near 56% across model generations 4. Real-world CVEs linked to AI tools are rising month over month 3. Enterprise data shows output and security findings growing faster than remediation 1. Together, these suggest that waiting for better models will not fix the problem.

One caveat matters. Veracode's test measures code written without an explicit request for secure output 2. That likely makes the results a floor rather than a ceiling, since careful prompting and review could improve them. But it also reflects how much "vibe coding" actually happens in practice.

The practical response is unglamorous:

• Treat AI output as untrusted input.
• Build static analysis and review into pipelines sized for AI-era commit volumes.
• Ask for security explicitly when prompting.

Compiling code was never the hard part. Teams that equate fluent output with safe output are the ones accumulating the debt.