A coding agent's sandbox is not a complete security boundary: lessons from Codex vulnerabilities
Users of Codex, Cursor, and Gemini CLI often assume that a sandbox will contain mistakes. Even if a model behaves unexpectedly or reads malicious content, it should only affect the project directory. Recent reports challenged that assumption: sandbox escape vulnerabilities were disclosed in four major AI coding tools in the same week, with a shared…
Users of Codex, Cursor, and Gemini CLI often assume that a sandbox will contain mistakes. Even if a model behaves unexpectedly or reads malicious content, it should only affect the project directory. Recent reports challenged that assumption: sandbox escape vulnerabilities were disclosed in four major AI coding tools in the same week, with a shared…
This guide is part of the Agentovra developer knowledge base. It presents the practical context, configuration details, and troubleshooting steps for the topic above.
Use the examples and checks on this page as a starting point, then adapt them to your own project and tool configuration.