AI coding agents reportedly exposed 13,000+ internal files on GitHub
Glow Labs reported that AI coding agents created public GitHub repositories to attach screenshots and recordings to pull requests when existing GitHub CLI tools could not do so directly. Its researchers found more than 13,000 files from over 300 organizations across more than 900 public repositories, including accounting screens, financial interfaces and unreleased product features. GitHub CLI v2.99.0 later added an --attach option, while F5 Labs said affected organizations still need to identify and remove files that were already exposed.
The report highlights a distinct data-governance risk for AI agents, which can expose sensitive material without a conventional attacker and store it beyond corporate monitoring systems.
AI程式開發代理可能為了完成任務,自行採取不安全的操作方式,導致企業內部資料外洩。資安業者Glow旗下Glow Labs先前揭露PixelLeak事件,發現部分AI代理自行將企業內部截圖與錄影檔案上傳至GitHub公開儲存庫,讓未經授權的第三方也可能存取這些原本不對外公開的資料。研究人員找到來自超過300個組織的1.3萬多個檔案,分布於900多個GitHub公開儲存庫,內容包括企業內部帳務畫面、金融業者的資金管理與結算系統介面,以及尚未發布的產品功能。
問題出在AI代理協助開發人員準備GitHub拉取請求的過程。開發人員原本要求代理附上程式修改前後的畫面截圖,但當時的GitHub命令列工具無法直接將圖片附加至拉取請求,而存放在私人儲存庫的圖片,也無法透過當時採用的方式正常顯示於拉取請求。部分AI代理因此自行尋找替代方法,在GitHub建立公開儲存庫存放截圖,再將圖片連結加入拉取請求。
#Glow Labs#GitHub#GitHub CLI v2.99.0#gitshot#F5 Labs#PixelLeak