Langflow RCE flaw CVE-2026-105740 draws 3,460 exploit attempts
iThome reports that a critical vulnerability in the open-source low-code AI development platform Langflow, tracked as CVE-2026-105740, has a severity score of 9.9 and has been targeted by more than 3,400 observed exploit attempts. The flaw lets authenticated attackers inject arbitrary system commands, including remote code execution, through MCP (Model Context Protocol) Stdio transport, and affects versions before 1.9.0; the maintainers published the advisory on the 5th of the month and released Langflow 1.9.0. Attack data comes from the security firm Previdian, which said it first saw attempted exploitation on September 19 and saw activity rise notably from October 11; by October 15 its honeypots had logged 1,423 attempts, and by the morning of October 22 the count reached 3,460 from 31 IP addresses across 17 countries and regions including Taiwan, China, Hong Kong, Japan, South Korea, Singapore, the US and Germany. Previdian said the CVE had not been added to CISA's Known Exploited Vulnerabilities catalog at the time of reporting.
Langflow is used to build AI applications and agents, so a 9.9-severity RCE with observed exploitation attempts means teams running older versions should apply the 1.9.0 update.
文/林妍溱|2026-10-22 發表
資安廠商揭露,低程式碼AI開發平臺Langflow一個風險值9.9的重大漏洞已觀測到超過3,400次漏洞利用嘗試。
Langflow是開源低程式碼開發平臺,讓開發人員能在視覺化介面運用整合工具開發AI應用及代理人。Langflow維護組織於本月5日公布CVE-2026-105740資安公告。本項漏洞可允許經過驗證的攻擊者透過MCP (Model Context Protocol)的Stdio傳輸功能注入任意系統指令,包括執行遠端程式碼。該漏洞影響1.9.0之前的版本。維護組織已發布Langflow 1.9.0更新版。
雖然CVE-2026-105740尚未被美國網路安全暨基礎設施安全局(CISA)列入已知受濫用漏洞(KEV)名冊,但資安業者Previdian根據第三方提供的證據以及該公司誘捕系統(Honeypot)感測器,判斷本項漏洞遭到濫用。
Previdian首先於9月19日觀測到其誘捕系統中該漏洞遭到試圖濫用,攻擊活動自10月11日起明顯增加。The Hacker News於10月15日報導時,Previdian誘捕系統偵測到針對CVE-2026-105740的試圖濫用活動為1,423次。截至10月22日上午,攻擊嘗試次數已增加至3,460次,攻擊流量來自31個IP位址,分布於臺灣、中國、香港、日本、韓國、新加坡、美國、德國等17個國家及地區。不確定這些IP流量所占比例,以及是否為伺服器被劫持或被當成中繼點。