AI-Generated Code Security Stalls as Codex and Claude Advance

Capability up, security flat

The newest coding models are better at almost everything except writing secure code. Recent security research and the current crop of flagship releases point in opposite directions. Vendors keep shipping more capable agents. Independent testing shows the security of the code those agents produce has barely changed.

On the capability side, OpenAI released GPT-5.5 on April 23, 2026. The company positions it as its new flagship, with stronger agentic coding, professional work and computer-use performance than GPT-5.4 4. The model now powers Codex, OpenAI's coding agent, on supported ChatGPT plans. Codex runs in the command line, IDE extensions, the web and iOS for repository-aware tasks 4. Anthropic, meanwhile, has made Claude Opus 4.7 its high-end option for difficult, long-running coding work. Sonnet stays the everyday engineering model, and Haiku handles cheaper automation 4. Opus 4.7 is also available through Replicate for teams that would rather not integrate directly with Anthropic's API 4.

What the security data shows

Veracode's 2026 GenAI Code Security Report tells a less flattering story. The company has tracked more than 100 large language models across four testing snapshots 2. It found an average security pass rate of 56%, up only slightly from 55% in its first report 23. Put the other way, roughly 44% of code-generation tasks produced code containing a known, exploitable vulnerability 23.

The test design matters. Veracode measured what happens when a developer does not explicitly ask for secure output 2. That is arguably the realistic default for most day-to-day prompting. The flaws involved are the kind a standard static or dynamic scanner would flag 2.

The Cloud Security Alliance's analysis frames the same research slightly differently. It says 45% of AI-generated samples introduce OWASP Top 10 vulnerabilities, and that the rate has not improved across testing cycles from 2025 into early 2026, despite vendor claims 1. The one-point gap between 44% and 45% probably reflects different snapshots or framing rather than a real disagreement. All three accounts agree on the core point: the trend line is flat.

Veracode's own summary describes the problem as "syntax is solved, security is not" 3. Its argument is that functional correctness and security are separate properties. Code that compiles, runs and looks tidy can still carry exploitable weaknesses, and that polish can give developers false confidence 3. The report also finds that greater model capability does not automatically bring better security 3.

Security debt at scale

The flat pass rate becomes a bigger problem because of how much code is now being generated. The Cloud Security Alliance cites research across Fortune 50 enterprises:

• AI-assisted developers commit code three to four times faster than their peers.
• They introduce security findings at ten times the rate.
• The result is a backlog of security debt that grows faster than organizations can fix it 1.

There are also early signs of real-world fallout. Georgia Tech's Vibe Security Radar project linked 35 CVEs in March 2026 alone directly to AI coding tools 1. Researchers estimate the true figure across open source is five to ten times higher 1. Attributing a CVE to a specific tool is hard, so these numbers should be read as indicative rather than precise. They still suggest the issue has moved beyond benchmarks.

Reading the gap

Taken together, the evidence leads to a clear conclusion. The industry's main metric for coding models is how much they can do on their own, and security is not improving along with it. GPT-5.5 and Opus 4.7 are marketed on agentic, long-running autonomy 4. More autonomy means more code reaching repositories with less line-by-line human review. If roughly two in five tasks still ship a known flaw by default, as Veracode reports 23, then faster and more independent agents will mostly produce vulnerable code at higher volume.

It would be unfair to blame the newest models specifically. None of the security research here benchmarks GPT-5.5 or Opus 4.7 by name, and these releases may perform differently. Still, Veracode's finding that capability gains have not translated into security gains 3 should make buyers skeptical of assuming they will.

The practical response for engineering teams is straightforward:

• Treat AI output as untrusted input.
• Ask for secure implementations explicitly, since the poor results are concentrated where developers don't 2.
• Keep scanners in the pipeline.
• Track remediation capacity against commit velocity, not just velocity alone.

The productivity gains are real. Without those controls, the security cost will keep growing alongside them.