LiteLLM says unrevoked key enabled poisoned 1.107.0 release
LiteLLM’s investigation, as reported by iThome, found that an Agent Platform master key exposed in a public GitHub configuration file in May was overwritten but never revoked. On October 10, an attacker used the key to obtain a GitHub organization owner’s CLI OAuth token, then stole two additional maintainers’ credentials through GitHub Actions workflows. The attacker redirected the release pipeline to the similarly named Berri-AI/litellm repository and published a compromised LiteLLM 1.107.0 containing credential-stealing code on PyPI and a container image registry; the affected release was removed on October 21.
LiteLLM’s account details how one unrevoked master key enabled credential theft and takeover of an official AI software release pipeline, making this a significant software supply-chain incident.
文/李建興|2026-10-26 發表
開源AI模型串接工具LiteLLM公布1.107.0版供應鏈攻擊的**調查結果**,發現攻擊者利用5月外洩後未撤銷的AI代理平臺主金鑰,取得GitHub組織擁有者的存取權杖。攻擊者隨後竊取另外兩名維護者的憑證,最終操控官方發布流程,讓含有竊密程式的版本上架PyPI及容器映像檔儲存庫。
5月LiteLLM團隊誤將實驗性Agent Platform的設定檔,提交至公開GitHub儲存庫,檔案內含平臺主金鑰。團隊雖然迅速透過強制推送覆寫該筆提交,也發現金鑰需要撤銷,卻未實際撤銷主金鑰。10月10日攻擊者使用該金鑰呼叫Agent Platform API,從回傳的AI代理設定資料中取得一名GitHub組織擁有者的CLI OAuth權杖。
取得權限後,攻擊者在專案儲存庫新增短暫使用的GitHub Actions工作流程,透過持續整合(CI)環境將機密資料傳送至外部端點。10月10日,兩個工作流程分別外洩另外兩名維護者的GitHub權杖,攻擊者隨即刪除執行紀錄,並於兩天後對其他私人儲存庫展開類似的竊密行動。