TLDR Tech / IT
Researchers escape OpenAI Codex sandbox to run commands on host
Researchers disclosed two flaws that allowed OpenAI Codex to escape its sandbox, including one that could execute commands on a developer's machine from the strictest read-only mode without an approval prompt. OpenAI fixed both issues, but the incident is another reminder that coding agents should be treated like privileged software: keep clients updated, limit local permissions, and be cautious about opening untrusted repositories with agentic tools.
Read the full story