AI Coding Agent Security: Sandbox Escapes & Fixes

Your AI coding assistant could be at risk. Researchers found seven sandbox escapes in tools like Cursor, OpenAI’s Codex, Google’s Gemini CLI, and Antigravity. These were triggered by hidden malicious instructions. Here's the thing: most of these issues are now patched. So, the first step is to update your software. Cursor users should move to version 3.0.0 or later. Codex users need version 0.95.0 or later. What's interesting is that updating alone isn't enough. The real trigger is prompt injection through untrusted content, often hidden in files like a README or a GitHub issue. Always be cautious with unfamiliar code. You should also scrutinize your dependencies. Pin them to known versions and avoid obscure packages. Limit the permissions you grant your coding agent and disable unused integrations. The bottom line: review proposed commands and edits before approving them. This simple pause can be your cheapest security upgrade.