Agent Sandbox Permissions & Least Privilege Matrix

Permission architecture matrix mapping read/write/execute capabilities for AI agents across local filesystems, cloud APIs, and shell environments.

Tags: security, permissions, least-privilege, sandbox, compliance

Category: devops

Author: SimplyUtils Team

Agent Sandbox Permissions & Least Privilege Matrix

Permission Tiers

Tier 1: Read-Only Exploration (Default / Low Risk)

Allowed: File reading, AST parsing, git status, git log, lint check execution.
Forbidden: File writing, network outbound calls, command execution with side effects.

Tier 2: Workspace Mutation (Medium Risk)

Allowed: Creating/modifying files within src/, tests/, docs/. Running unit test suites.
Guardrails: Excludes .env*, .git/config, id_rsa, and OS system directories.

Tier 3: External Integration & Terminal Execution (High Risk)

Allowed: Running npm install, database migrations, making external API requests.
Mandatory Control: Explicit human approval dialog before execution.