Security researchers have demonstrated two methods for escaping OpenAI's Codex sandbox. In one of the attacks, they were able to run commands on a developer's machine even when the sandbox was in its most locked-down mode.
Both vulnerabilities have now been patched by OpenAI. The findings highlight the difficulty of containing AI coding agents, which are designed to execute code and interact with external systems.
This report is based on a single source, so there are no conflicting accounts to compare. The key takeaway is that both escape techniques were fixed before the details became public.