AI Agent Safety for JavaScript and Node.js Developers

SafeClaw by Authensor provides deny-by-default action gating for JavaScript and Node.js AI agents. It intercepts child_process.exec(), fs.writeFile(), fetch(), and any other dangerous operation, checking each against your YAML policy before allowing execution. Install with npx @authensor/safeclaw — zero dependencies, works with Claude and OpenAI agents alike.

The Node.js Agent Risk Surface

Node.js agents have unrestricted access to child_process, fs, net, and http. A single compromised tool call can:

• Execute rm -rf / via child_process.exec()
• Read .env files containing API keys via fs.readFileSync()
• Exfiltrate data via fetch() or axios.post() to an attacker-controlled endpoint
• Install malicious packages via child_process.exec("npm install trojan-package")

Installation

npx @authensor/safeclaw

This starts SafeClaw as a local sidecar. Your Node.js agent communicates with it before executing any action.

Policy Configuration

Create safeclaw.policy.yaml:

version: 1
defaultAction: deny

rules:
  - action: file.read
    path:
      glob: "./src/**"
    decision: allow

- action: file.write
    path:
      glob: "./output/**"
    decision: allow

- action: file.read
    path:
      glob: "*/.env"
    decision: deny  # never let agents read env files

- action: process.exec
    command:
      startsWith: "node"
    decision: allow

- action: process.exec
    command:
      startsWith: "npm install"
    decision: prompt

- action: network.request
    host:
      in: ["api.openai.com", "api.anthropic.com"]
    decision: allow

JavaScript Integration

import { Gate } from '@authensor/safeclaw';
import { execSync } from 'child_process';
import fs from 'fs/promises';

const gate = new Gate();

async function safeExec(command) {
  const decision = await gate.check({
    action: 'process.exec',
    command
  });
  if (!decision.allowed) {
    throw new Error(SafeClaw denied: ${decision.reason});
  }
  return execSync(command, { encoding: 'utf-8' });
}

async function safeReadFile(filepath) {
  const decision = await gate.check({
    action: 'file.read',
    path: filepath
  });
  if (!decision.allowed) {
    throw new Error(SafeClaw denied: ${decision.reason});
  }
  return fs.readFile(filepath, 'utf-8');
}

async function safeFetch(url, options = {}) {
  const host = new URL(url).hostname;
  const decision = await gate.check({
    action: 'network.request',
    host,
    url,
    method: options.method || 'GET'
  });
  if (!decision.allowed) {
    throw new Error(SafeClaw denied: ${decision.reason});
  }
  return fetch(url, options);
}

Wrapping Axios

If your agent uses axios for HTTP:

import axios from 'axios';

async function safeAxios(config) {
  const url = typeof config === 'string' ? config : config.url;
  const host = new URL(url).hostname;
  const decision = await gate.check({
    action: 'network.request',
    host,
    url,
    method: config.method || 'GET'
  });
  if (!decision.allowed) {
    throw new Error(SafeClaw denied: ${decision.reason});
  }
  return axios(config);
}

Gating npm install

async function safeNpmInstall(packageName) {
  const decision = await gate.check({
    action: 'process.exec',
    command: npm install ${packageName}
  });
  if (!decision.allowed) {
    throw new Error(SafeClaw denied npm install ${packageName}: ${decision.reason});
  }
  return execSync(npm install ${packageName}, { encoding: 'utf-8' });
}

With the prompt decision in your policy, this pauses and asks a human before allowing the install.

Audit Trail

Every decision is hash-chained:

{
  "timestamp": "2026-02-13T10:30:00Z",
  "action": "file.read",
  "path": "./.env",
  "decision": "deny",
  "reason": "matched deny rule for .env files",
  "hash": "b4e9f2...",
  "prev_hash": "a3f8c1..."
}

MIT licensed. Works with Claude Agent SDK, OpenAI Assistants, Vercel AI SDK, and any custom agent framework.

Cross-References

• Next.js Integration
• Express.js Integration
• TypeScript Integration
• Deny-by-Default Explained
• Prevent Agent .env File Access