OpenAI announced an upgrade to the security system of its AI coding agent Codex, implementing multiple security measures to reduce the risk of potential destructive operations such as accidentally deleting files or overwriting data during code execution tasks.

Recently, OpenAI team member Tibo posted on a social platform stating that the company investigated reports of destructive operations from Codex and found that the main risks were concentrated in the temporary working directory cleanup phase. Some issues stemmed from the reuse of system variables, such as incorrectly using variables like $HOME, which could cause cleanup operations to point to the actual user directory; additionally, some operations lacked sufficient path verification before deletion or overwriting, increasing the risk of data damage.

To address these issues, OpenAI has enhanced protection at multiple levels. The new version of Codex requires agents to verify the target path before performing sensitive operations and uses independent temporary directories, avoiding reliance on potentially misleading system variables. Meanwhile, OpenAI has strengthened the identification and interception mechanisms for dangerous commands and added related operation review processes.

In addition, OpenAI has raised the threshold for using the Codex Full access permission, lowering the possibility of users mistakenly granting authorization through increased risk warnings, and upgraded Auto-review rules, further controlling security risks while maintaining AI programming efficiency.

As AI coding agents gradually participate in code development, file management, and automated task execution, the safety of agent behavior has become a key focus for the industry. This upgrade by OpenAI highlights the need for AI Agents to simultaneously improve permission management, operation auditing, and risk control capabilities as their application scope expands.