Science & Tech · AI · 3 hrs ago
Security in AI-generated code improves, but complex flaws remain
A study found security in AI-generated code improved for the first time in a year.
Researchers tested 18 models in four programming languages.
The share of code with vulnerabilities fell from 44% to 38%, while the overall security score rose from 56% to 62%.
The gains mainly involve common, pattern-based flaws, such as cross-site scripting.
The leading models scored 73%, but they still struggle with complex security reasoning across multiple functions.
One expert said even the most expensive models fail on more than one in four security tasks.
Organizations are urged to test AI-generated code independently and keep verifying it before deployment, with human oversight still needed.
Security in code generated by frontier AI models improved for the first time in a year, with vulnerability rates falling from 44% to 38%.
A Veracode report found that the overall security score for AI-generated code rose from 56% to 62%.
Researchers analyzed 18 models across four programming languages.
The leading models named in the report, Claude Opus 5.5 and GPT-6-Sol, achieved a 73% security rating.
The models still struggle with complex security reasoning, and researchers say organizations should use independent testing and human oversight.
- Who
- Researchers analyzed AI models, and Veracode reported the findings.
- What
- Security scores for AI-generated code rose, but complex flaws remain difficult for models to prevent.
- When
- The report was covered on October 30, 2026.
- Where
- Not stated.
- Why
- Not stated.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Vulnerability rate
- Decreased from 44% to 38%
- Overall security score
- Rose from 56% to 62%
- Models analyzed
- 18
- Programming languages
- 4
- Leading models' security rating
- 73% for Claude Opus 5.5 and GPT-6-Sol