sandbox-hardening

What problem does it solve?

Sandbox Hardening prevents AI agents from causing damage when they run generated or untrusted code by isolating execution, limiting resources, and restricting network and filesystem access.

Core Features & Use Cases

• Isolation level selection: Choose process-level, container-level, or VM-level isolation based on trust boundary (internal tooling vs multi-tenant vs untrusted code).
• Least-privilege enforcement: Apply permission boundaries, drop capabilities, run as non-root, and scope filesystem mounts to reduce what the agent can touch.
• Resource and network control: Set CPU, memory, PID, and file descriptor limits while using network allowlists and syscall auditing to limit blast radius.
• Validation-before-execution: Reduce risk by validating commands and applying blocklists to prevent obviously dangerous operations.

Quick Start

Instruct your agent to run its generated code inside a least-privilege sandbox using a container configured with a read-only filesystem, dropped Linux capabilities, strict resource limits, and an allowlisted network.