NVIDIA's AI Red Team outlines mandatory and recommended security controls for sandboxing AI coding agents to mitigate indirect prompt injection attacks. Mandatory controls include blocking network egress to arbitrary sites, preventing file writes outside the workspace, and blocking writes to configuration files. Recommended measures include sandboxing the entire IDE, using virtualization to isolate the kernel, preventing reads outside the workspace, requiring manual approval for isolation violations, implementing secret injection, and establishing sandbox lifecycle management. The guidance emphasizes OS-level enforcement over application-level controls since agents execute arbitrary code by design.
Table of contents
Why enforce sandbox controls at an OS level?Mandatory sandbox security controlsRecommended sandbox security controlsLearn more140 Impressions