According to Tech Radar, security in code generated by frontier artificial intelligence models has improved for the first time in a year, with vulnerability rates decreasing from 44% to 38%.
A new report from Veracode indicates that while large language models are becoming more proficient at preventing common flaws like cross-site scripting, complex security reasoning remains a challenge. Researchers analyzed 18 models across four programming languages and found that the overall security score for AI-generated code rose from 56% to 62%. Leading models, including Claude Opus 5.5 and GPT-6-Sol, achieved a 73% security rating. Despite these gains, experts warn that the improvement is largely driven by the models' ability to address simple, pattern-based vulnerabilities. When tasks require complex data-flow reasoning across multiple functions, such as preventing log injection, the models continue to struggle. Chris Wysopal, co-founder and chief security evangelist at Veracode, emphasized that even the most expensive models fail on more than one in four security tasks. Consequently, researchers stress that organizations must not rely solely on AI output. Independent security testing and continuous verification remain essential to ensure code integrity before deployment, as AI-generated code still requires human oversight to mitigate persistent security risks.
Source: Tech Radar
