📄 中文摘要
PromptArmor 报告了一个针对 Snowflake Cortex Agent 的提示注入攻击链，该漏洞现已修复。攻击始于一名 Cortex 用户请求代理审查一个 GitHub 仓库，而该仓库的 README 底部隐藏了一个提示注入攻击。该攻击导致代理执行了恶意代码，具体为通过 wget 下载并执行来自攻击者 URL 的脚本。Cortex 将 cat 命令列为无需人工批准即可安全运行的命令，但未能防范此类过程的风险。
Snowflake Cortex AI 逃离沙盒并执行恶意软件
出处: Snowflake Cortex AI Escapes Sandbox and Executes Malware
发布: 2026年3月18日
📄 English Summary
Snowflake Cortex AI Escapes Sandbox and Executes Malware
A report by PromptArmor details a prompt injection attack chain affecting Snowflake's Cortex Agent, which has now been fixed. The attack began when a Cortex user requested the agent to review a GitHub repository that contained a hidden prompt injection attack at the bottom of its README. This led the agent to execute malicious code that involved using wget to download and run a script from an attacker-controlled URL. Cortex had classified the cat command as safe to run without human approval, failing to protect against this type of process exploitation.