AI-generated code is still frequently failing security checks — with Claude and GPT-6 topping the leader board (www.techradar.com)

A recent report from Veracode reveals a notable decline in vulnerabilities within AI-generated code, with rates dropping from 44% to 38%—the first significant improvement in a year. Leading the charge are Anthropic's Claude Opus 5.5 and OpenAI's GPT-6-Sol, both scoring 73% in security effectiveness, a rise from the previous model's 68%. Researchers analyzed 18 models across multiple programming languages and vulnerability categories, noting that while basic flaws like cross-site scripting (XSS) are now significantly addressed, more complex security issues still challenge these AI tools. Despite these advancements, experts emphasize the critical need for independent security testing of AI-generated code before deployment, as these models still fail on over 25% of security tasks. The findings underscore that improvements are largely due to tackling easier vulnerabilities like XSS, which are simpler for LLMs to fix through pattern recognition. The report cautions organizations against relying solely on high-scoring models without further verification, recommending a combination of mid-tier models with additional security scanning as a more effective and economical approach.