Agent Sandbox

Agent Sandbox

Overview

AI agents execute code, modify files, and run shell commands. Without guardrails, a bad prompt or hallucination can delete your database, overwrite production configs, or exfiltrate secrets. This skill builds safety layers — sandboxed execution, filesystem restrictions, network policies, audit trails, and kill switches.

When to Use

• Running untrusted or AI-generated code in production
• Adding safety controls to coding agents that modify your codebase
• Restricting which files, directories, or commands an agent can access
• Logging every agent action for compliance or debugging
• Building multi-tenant agent platforms where agents need isolation

Instructions

Strategy 1: Filesystem + Process Sandbox (Zero Dependencies)

The simplest safety layer — restrict which paths the agent can read/write and which commands it can execute. No Docker required.

// sandbox.ts — Filesystem and process sandbox for AI agents
/**
 * Wraps agent operations with safety checks:
 * - Allowlist/denylist for file paths
 * - Command blocklist (rm -rf, DROP TABLE, etc.)
 * - Audit log of every action
 * - Kill switch to halt agent immediately
 */
import { execSync } from "child_process";
import { readFileSync, writeFileSync, existsSync, appendFileSync } from "fs";
import { resolve, relative } from "path";

interface SandboxConfig {
  workDir: string;                    // Root directory agent can access
  allowedPaths: string[];             // Glob patterns of allowed paths
  deniedPaths: string[];              // Glob patterns of denied paths
  blockedCommands: string[];          // Commands that are never allowed
  maxFileSize: number;                // Max bytes per file write
  auditLog: string;                   // Path to audit log file
  readOnly: boolean;                  // If true, block all writes
}

const DEFAULT_BLOCKED = [
  "rm -rf /", "rm -rf ~", "rm -rf .",
  "mkfs", "dd if=", "> /dev/sd",
  "DROP DATABASE", "DROP TABLE", "TRUNCATE",
  "curl.*|.*sh", "wget.*|.*bash",           // Pipe to shell
  "chmod 777", "chmod -R 777",
  "env | curl", "printenv | curl",           // Secret exfiltration
  "ssh-keygen", "ssh-copy-id",
];

export class AgentSandbox {
  private config: SandboxConfig;
  private killed = false;

  constructor(config: Partial<SandboxConfig> & { workDir: string }) {
    this.config = {
      allowedPaths: ["**"],
      deniedPaths: ["**/.env", "**/.ssh/**", "**/node_modules/**"],
      blockedCommands: DEFAULT_BLOCKED,
      maxFileSize: 1024 * 1024,  // 1MB default
      auditLog: "./agent-audit.jsonl",
      readOnly: false,
      ...config,
    };
  }

  /**
   * Read a file through the sandbox — checks path is allowed.
   */
  readFile(filePath: string): string {
    this.checkKilled();
    const absPath = resolve(this.config.workDir, filePath);
    this.checkPathAllowed(absPath, "read");
    this.audit("read", filePath);
    return readFileSync(absPath, "utf-8");
  }

  /**
   * Write a file through the sandbox — checks path, size, and read-only mode.
   */
  writeFile(filePath: string, content: string): void {
    this.checkKilled();
    if (this.config.readOnly) {
      throw new SandboxError("Write blocked: sandbox is read-only");
    }

    const absPath = resolve(this.config.workDir, filePath);
    this.checkPathAllowed(absPath, "write");

    if (Buffer.byteLength(content) > this.config.maxFileSize) {
      throw new SandboxError(
        `Write blocked: file exceeds max size (${this.config.maxFileSize} bytes)`
      );
    }

    this.audit("write", filePath, { size: Buffer.byteLength(content) });
    writeFileSync(absPath, content);
  }

  /**
   * Execute a command through the sandbox — checks against blocklist.
   */
  exec(command: string, timeoutMs: number = 30000): string {
    this.checkKilled();
    this.checkCommandAllowed(command);
    this.audit("exec", command);

    try {
      return execSync(command, {
        cwd: this.config.workDir,
        encoding: "utf-8",
        timeout: timeoutMs,
        maxBuffer: 10 * 1024 * 1024,  // 10MB output limit
      });
    } catch (error: any) {
      this.audit("exec_error", command, { error: error.message });
      throw error;
    }
  }

  /**
   * Kill switch — immediately halt all agent operations.
   */
  kill(reason: string): void {
    this.killed = true;
    this.audit("killed", reason);
    console.error(`🛑 Agent sandbox killed: ${reason}`);
  }

  private checkKilled(): void {
    if (this.killed) throw new SandboxError("Agent has been killed");
  }

  private checkPathAllowed(absPath: string, operation: string): void {
    const relPath = relative(this.config.workDir, absPath);

    // Must be within workDir (no ../ escapes)
    if (relPath.startsWith("..")) {
      throw new SandboxError(`${operation} blocked: path escapes sandbox (${relPath})`);
    }

    // Check denylist
    for (const pattern of this.config.deniedPaths) {
      if (matchGlob(relPath, pattern)) {
        throw new SandboxError(`${operation} blocked: path matches denylist (${pattern})`);
      }
    }
  }

  private checkCommandAllowed(command: string): void {
    const lower = command.toLowerCase();
    for (const blocked of this.config.blockedCommands) {
      if (lower.includes(blocked.toLowerCase())) {
        throw new SandboxError(`Command blocked: matches "${blocked}"`);
      }
    }
  }

  private audit(action: string, target: string, extra?: Record<string, unknown>): void {
    const entry = {
      timestamp: new Date().toISOString(),
      action,
      target,
      ...extra,
    };
    appendFileSync(this.config.auditLog, JSON.stringify(entry) + "\n");
  }
}

class SandboxError extends Error {
  constructor(message: string) {
    super(message);
    this.name = "SandboxError";
  }
}

function matchGlob(path: string, pattern: string): boolean {
  const regex = pattern
    .replace(/\*\*/g, ".*")
    .replace(/\*/g, "[^/]*")
    .replace(/\?/g, ".");
  return new RegExp(`^${regex}$`).test(path);
}

Strategy 2: Docker Container Sandbox

For full isolation — the agent runs inside a container with limited CPU, memory, network, and filesystem access.

// docker-sandbox.ts — Run agent code in isolated Docker containers
/**
 * Spawns a Docker container for each agent session.
 * Mounts only the allowed workspace directory (read-only optional).
 * Enforces CPU, memory, and network limits.
 * Auto-kills containers that exceed time limits.
 */
import { execSync, spawn } from "child_process";

interface DockerSandboxConfig {
  image: string;                 // Base image (e.g., "node:20-slim")
  workDir: string;               // Host directory to mount
  readOnly: boolean;             // Mount workspace as read-only
  cpuLimit: string;              // CPU limit (e.g., "1.0" = 1 core)
  memoryLimit: string;           // Memory limit (e.g., "512m")
  networkMode: string;           // "none" for no network, "bridge" for limited
  timeoutSeconds: number;        // Kill container after this many seconds
  allowedEnvVars: string[];      // Env vars to pass through
}

export class DockerSandbox {
  private config: DockerSandboxConfig;
  private containerId: string | null = null;

  constructor(config: Partial<DockerSandboxConfig> & { workDir: string }) {
    this.config = {
      image: "node:20-slim",
      readOnly: false,
      cpuLimit: "1.0",
      memoryLimit: "512m",
      networkMode: "none",        // No network by default
      timeoutSeconds: 300,        // 5 minutes max
      allowedEnvVars: [],
      ...config,
    };
  }

  /**
   * Start the sandbox container.
   */
  async start(): Promise<string> {
    const mountFlag = this.config.readOnly ? "ro" : "rw";
    const envFlags = this.config.allowedEnvVars
      .map((v) => `-e ${v}`)
      .join(" ");

    const cmd = [
      "docker run -d",
      `--cpus=${this.config.cpuLimit}`,
      `--memory=${this.config.memoryLimit}`,
      `--network=${this.config.networkMode}`,
      "--security-opt=no-new-privileges",    // No privilege escalation
      "--read-only",                          // Root FS read-only
      "--tmpfs /tmp:size=100m",              // Writable tmp with size limit
      `-v ${this.config.workDir}:/workspace:${mountFlag}`,
      `-w /workspace`,
      envFlags,
      this.config.image,
      "tail -f /dev/null",                   // Keep container alive
    ].join(" ");

    this.containerId = execSync(cmd, { encoding: "utf-8" }).trim();

    // Auto-kill timer
    setTimeout(() => this.kill("timeout"), this.config.timeoutSeconds * 1000);

    return this.containerId;
  }

  /**
   * Execute a command inside the sandbox container.
   */
  exec(command: string): string {
    if (!this.containerId) throw new Error("Sandbox not started");
    return execSync(
      `docker exec ${this.containerId} sh -c '${command.replace(/'/g, "'''")}'`,
      { encoding: "utf-8", timeout: 60000 }
    );
  }

  /**
   * Kill the sandbox container and remove it.
   */
  kill(reason: string = "manual"): void {
    if (this.containerId) {
      console.log(`🛑 Killing sandbox: ${reason}`);
      execSync(`docker kill ${this.containerId} && docker rm ${this.containerId}`, {
        encoding: "utf-8",
      });
      this.containerId = null;
    }
  }
}

Examples

Example 1: Add safety controls to a coding agent

User prompt: "I want my AI coding agent to only modify files in the src/ directory and never touch .env files or run destructive commands."

The agent will:

• Create an AgentSandbox with workDir pointing to the project root
• Set allowedPaths to ["src/**", "tests/**"]
• Set deniedPaths to ["**/.env*", "**/.ssh/**", "**/secrets/**"]
• Enable audit logging to track every agent action
• Wrap all file operations and command executions through the sandbox

Example 2: Run untrusted code in Docker isolation

User prompt: "We're building a code execution platform. User-submitted code needs to run in isolation with no network access, 512MB memory, and a 30-second timeout."

The agent will:

• Set up DockerSandbox with networkMode "none", memory 512m, timeout 30s
• Mount user code directory as read-only
• Enable writable /tmp with 100MB limit for temporary files
• Add security-opt no-new-privileges to prevent escalation
• Implement cleanup on timeout or completion

Guidelines

• Default to deny — block everything, then allowlist what the agent needs
• No Docker socket access — mounting docker.sock gives root on the host
• Audit everything — log every file read, write, and command for forensics
• Time limits prevent infinite loops — always set exec timeouts
• Network "none" by default — agents shouldn't make outbound calls unless explicitly needed
• Read-only mounts when possible — agents that only analyze code don't need write access
• Separate audit logs from agent workspace — the agent shouldn't be able to modify its own audit trail
• Test the sandbox itself — try to escape it before trusting it with real data
• Kill switches save you — always have a way to halt the agent immediately
• Container cleanup — always docker rm after container stops to avoid disk waste