Researchers escape OpenAI Codex sandbox to run commands on host

Security researchers have successfully bypassed the sandbox environment of OpenAI's Codex in two different ways, demonstrating potential security gaps in AI deployment.

Notably, one of the methods allowed executing commands directly on a developer's machine even while operating in its most locked-down and secure mode.

According to BleepingComputer, OpenAI was quick to respond and has already patched both vulnerabilities to ensure user safety and system integrity.

This discovery highlights the ongoing challenges in securing AI-driven tools that interact with local development environments and execute generated code.

For the broader tech community, this serves as an important reminder of the critical need for rigorous security audits when integrating advanced AI assistants into daily workflows.