Technology RadarTechnology Radar

AI Agent Sandboxing

ai-securitycontainers
Adopt

AI agent sandboxing uses OS-level primitives (Landlock, seccomp, Seatbelt) and/or virtualization (containers, MicroVMs) to constrain what AI coding agents can do. Every major agent now ships with some form of sandboxing — OpenAI Codex uses Landlock + seccomp by default, Cursor uses Seatbelt/Landlock, and Claude Code uses filesystem + network isolation with seccomp BPF.

How Major Agents Handle It

Agent Sandboxing Default Mode
OpenAI Codex CLI Landlock + seccomp Workspace-write (network disabled)
Cursor Seatbelt (macOS) / Landlock (Linux) Dynamic policy per workspace
Claude Code Filesystem + network isolation + seccomp BPF Specific directory access only
E2B Firecracker MicroVMs Per-execution VM (~150ms cold start)

Key Technologies

  • Landlock LSM (kernel 5.13+): Capability-based filesystem restrictions below the application layer — agents cannot override them
  • seccomp-BPF: System call filtering that blocks network-related and dangerous syscalls
  • landrun (2K+ GitHub stars): Most popular standalone Landlock wrapper, no root or containers needed
  • jai (Stanford): Copy-on-write overlay sandbox — one command, no root, no containers. Gives the working directory full access and hides the rest of $HOME behind an overlay
  • Firecracker MicroVMs: Dedicated guest kernel, ~150ms cold start, strongest isolation (E2B scaled to 15M sandboxes/month)

Strengths

  • OS-level enforcement cannot be bypassed by the agent
  • Multiple approaches for different threat models (kernel-level vs. containers vs. MicroVMs)
  • Measurable impact — Claude Code reduced permission prompts by 84%, Cursor reduced interruptions ~40%

Limitations

  • Landlock/seccomp are Linux-specific
  • No standardized approach across agents
  • GPU passthrough is challenging in sandboxed environments
  • Kernel-level sandboxing is lighter weight but less isolated than MicroVMs