Researchers escape OpenAI Codex sandbox to run commands on host

Refract AI Intelligence Digest

BLUF

OpenAI patched critical sandbox escapes in Codex that allowed host machine compromise.

NEWS

Researchers demonstrated two distinct methods to bypass OpenAI's Codex security sandbox, including one that executed commands on a developer's local machine from the most restricted mode. Both vulnerabilities have been identified and patched by OpenAI following responsible disclosure.

Why I Care

This highlights risks in AI-assisted coding tools where malicious prompts could lead to unauthorized system access. Developers and enterprises using Codex for sensitive codebases face potential data exfiltration or malware installation if unpatched versions were used.

Next Steps

Verify your Codex environment is updated to the latest version immediately. Review recent code generation logs for anomalies if you use unpatched instances.

Researchers escaped OpenAI's Codex sandbox two ways, one running commands on a developer's machine from its most locked-down mode. OpenAI has patched both. [...]
Back to Blog Listing

Source: BleepingComputer ·

This digest was generated by Refract AI Collective to help the public sector security community stay informed.