AI Coding Agents Face New Tests and Accountability Questions
TechNews reports that AI-generated code can run faster yet still conflict with system architecture, duplicate existing features, and create maintenance risks. XData Group CEO Roman Eloshvili cited Veracode research that found known vulnerabilities in roughly 44% of AI-written code tasks tested in 2026, while recommending review by a separate model to avoid repeating the same blind spots. Anthropic data relayed by TechNews indicates that, across about 400,000 Claude Code sessions, humans usually planned tasks while AI executed them. Cequence Security CTO Shreyans Mehta warned that rapid agent operations can turn human-in-the-loop review into a rubber stamp, and proposed shared accountability among model developers, deploying enterprises, and the industry ecosystem.
The report shifts attention from AI coding speed toward software quality, reviewability, and allocation of responsibility for failures in autonomous development workflows.


生成式 AI 大幅加速程式開發,AI 代理(AI agent)也逐步自動執行任務,但出了問題該由誰負責,成為產業必須面對的課題。兩位科技業高層近日在《富比士》(Forbes)科技委員會專欄撰文,分別從程式品質與管理 AI 代理切入,指出責任不會因自動化消失,而是轉移到掌握控制權的一方。
專注歐洲銀行業的B2B軟體開發商XData Group創辦人暨執行長Roman Eloshvili指出,AI寫程式的速度遠快於人工,但會寫程式不等於交出可靠的軟體。AI程式單獨看或許正確,卻可能不符整體系統架構,如重複既有功能,埋下日後維護隱憂。他引述Veracode研究,2026年測試約44%的AI寫程式任務含已知漏洞。
#XData Group#Veracode#Anthropic#Claude Code#Cequence Security