31 minutes ago
3
- AI-generated code security improved for the first time in a year, with vulnerability rates dropping from 44% to 38%
- Top models performed better at preventing common flaws like XSS, but still struggled with complex security reasoning
- Researchers stress AI-generated code still requires independent security testing and verification before deployment
Computer code generated by frontier AI models has gotten more secure and consequently - contains fewer vulnerabilities.
This is according to a new report by Veracode, which claims this is the first significant improvement in a year.
In July 2025, Veracode started releasing its GenAI Code Security Report, which analyzed code generated by LLMs for vulnerabilities and security risks. Since then, the company’s researchers noticed a trend - the code was functioning almost perfect, but on the security side it left much to be desired.
Are XSS flaws worth the trouble?
“We’ve been analyzing code generated by multiple LLMs for more than a year. Until now, we have not seen a marked improvement in AI code security, despite consistently near-perfect functionality,” said Chris Wysopal, Co-founder & Chief Security Evangelist at Veracode.
“This is the first time since running our study that we’ve been able to report a meaningful move on the security dial.”
That meaningful move is a jump from 56% to 62% in overall AI-generated code security. At the same time, there’s been a drop in vulnerabilities from 44% to 38%, Veracode added, saying this was the most significant improvement since the reports started coming out. Anthropic's Claude Opus 5.5 and OpenAI's GPT-6-Sol co-lead the rankings at 73%, surpassing the previous high of 68% set by GPT-5.5 in July 2026.
“Ten models now clear 60%, up from four in the last report, and the overall average has jumped six percentage points,” Wysopal said. “With around half of code now generated by AI, today’s results are truly encouraging.”
To draft the report, Veracode tested 18 models across four languages (Python, C#, JavaScript, and Java), and four Common Weakness Enumeration (CWE) categories (CWE-89, CWE-327, CWE-80, CWE-117). Claude Opus 5.5 and GPT-6-Sol claimed the top spot with 73% followed by GPT-5.5 which holds 68%. In July, six of 11 models scored between 50 and 53, while today only two remain in that band, while 10 scored at least 60%. The bottom is still at 44% - 50% percent.
While the results are definitely commendable, there is a caveat. Apparently, the main driver of the overall gain was on cross-site scripting (XSS), which is arguably one of the weakest categories. XSS rose to 18% which is also the single biggest change in the dataset. According to Veracode, XSS is the “easiest vulnerability to fix through pattern recognition”, which is exactly how LLMs improve. In cases where data-flow reasoning across multiple functions is required (such as, in log injection cases, for example), LLMs continue to stall.
Low-hanging fruit
In other words, AI has just picked low-hanging fruit and called it a day. We’re not sure if we should be disappointed with this, since XSS attacks are common and the vulnerability is still present across the web.
In the report, Veracode also said that water was wet and the sky is mostly blue: it found that top-performing models cost more.
However, they said throwing money at the problem most likely isn’t the best solution: “The models that produce the most secure code are the ones that cost the most to run at scale, yet they still fail on more than one in four security tasks,” Wysopal said. “The answer should never be to use the most expensive model and hope for the best. A mid-tier model at a 60% pass rate plus Veracode scanning is more accurate, more cost-effective, and more scalable than a frontier model at 73% with no verification.”
In conclusion, the improvement is there, but “mends a small part of a larger wound”, Veracode said. Organizations should make sure they combine AI code with continuous and independent security verification that can catch what the models miss.
