Skip to content

Agent Watchdog

AI agents like Cursor, Claude, and Copilot now run shell commands, read files, and even make payments through agentic wallets.

Code Video Cerebral Valley

Core trick

Uses eBPF process monitoring to block agent-launched commands by default, with passkey-approved scoped grants for specific tasks.

Tags

securityagents

Description

AI agents like Cursor, Claude, and Copilot now run shell commands, read files, and even make payments through agentic wallets. They have real-world power without a permission system. A prompt injection or buggy plugin can leak API keys, run dangerous CLI commands, or spend money without the owner’s knowledge. Today the only defense is “don’t give agents access,” which blocks real productivity gains. We built Agent Watchdog to solve this. It uses system-level process monitoring to block risky actions by default, lets owners grant scoped access with passkey authentication for specific tasks, budgets, and time limits, and provides a dashboard to audit every action and payment. Agents stay autonomous within scope, but the owner always keeps control.
not-so-fat/agent_watchdog★ 00 forksTypeScriptCreated Feb 28, 2026 · last push Mar 1, 2026

README excerpt

# Agent Watchdog -- Hackathon Demo

An **Agent Watchdog** that monitors what AI agents do on your machine and controls how they spend your money. **Watchdog** is an eBPF-based runtime that can block or allow processes (e.g. which PIDs may execute payment CLIs). **Pulse** provides scoped access (grants), policy, and x402 payment; the user approves grants via the dashboard. The demo uses **Cursor as the AI agent**: Cursor is not allowed to run `awal` directly; it must go through Pulse.

## System Overview

```
                         User (browser)
                  Dashboard + Approval UI
                  http://localhost:4020/dashboard
                           |
      +--------------------+---------------------+
      |                                           |
+-----v------+                          +---------v---------+
|  Watchdog  |                          |  Pulse            |
|  (eBPF)    |                          |  (Node.js :4020)  |
|            |                          |                   |
|  Enforce   |                          |  Scoped Access    |
|  which PIDs|                          |  Policy engine    |
|  may run   |                          |  x402 engine      |
|  awal      |                          |  Dashboard        |
+-----+------+                          +---------+---------+
      |  monitors                                 |  calls
      v                                           v
+-------------+                          +-----------

Team