OpenAI Codex 存在命令注入漏洞，可能窃取您的 GitHub 令牌

BeyondTrust's Phantom Labs recently published a report detailing a command injection vulnerability in OpenAI's Codex. Although the vulnerability has been patched, the attack pattern is significant as it represents a security risk that many developers may not anticipate. Codex operates tasks within managed containers that clone users' GitHub repositories and authenticate using short-lived OAuth tokens. The vulnerability arose because branch names were not sanitized before being passed to shell commands during environment setup. An attacker could craft a malicious branch name that injects arbitrary shell commands, which would execute within the container and have access to the user's GitHub token. This attack method was applicable across multiple interfaces of Codex, including the web interface, CLI, SDK, and IDE integrations.