OpenAI patched two reported exploits that allowed code to escape the Codex sandbox and run commands outside it
Security issue Provisional 78% confidence first seen
Two separate sandbox-escape vulnerabilities in OpenAI Codex were reported on August 12, 2026, including one in the Codex CLI patch flow and another in the Codex Desktop JavaScript tool. OpenAI fixed both issues within eight days by updating Codex to close the identified sandbox-permission and cross-context execution paths, preventing the previously described unsandboxed command execution.