Security researchers found two ways out of the OpenAI Codex sandbox, one of them capable of running commands on a developer's machine from Codex's most locked-down mode, with no approval prompt and ...