How to Stop AI Agents from Opening Network Ports

SafeClaw by Authensor prevents AI agents from opening network ports, starting HTTP servers, or binding to any network interface through deny-by-default action gating. Any command that would create a listening socket is denied unless you explicitly allow it. Install with npx @authensor/safeclaw and unauthorized network listeners are blocked from the first agent action.

Why Port Opening Is Dangerous

When an AI agent opens a network port, it creates an entry point into your machine. This could allow remote access by attackers, expose internal services to the public internet, create a reverse shell for persistent access, or conflict with existing services on the same port. An agent running python -m http.server 8080 in your home directory exposes every file to anyone on your network.

Step 1: Install SafeClaw

npx @authensor/safeclaw

Zero dependencies, MIT licensed. Works with Claude, OpenAI, and all agent frameworks.

Step 2: Block Server-Starting Commands

# safeclaw.policy.yaml
rules:
  # Block common server commands
  - action: shell.execute
    command_pattern: "python -m http.server"
    effect: deny
    reason: "Block Python HTTP server"

- action: shell.execute
    command_pattern: "python -m SimpleHTTPServer"
    effect: deny
    reason: "Block Python 2 HTTP server"

- action: shell.execute
    command_pattern: "npx serve*"
    effect: deny
    reason: "Block npx serve"

- action: shell.execute
    command_pattern: "npx http-server*"
    effect: deny
    reason: "Block npx http-server"

- action: shell.execute
    command_pattern: "php -S *"
    effect: deny
    reason: "Block PHP built-in server"

- action: shell.execute
    command_pattern: "ruby -run -e httpd*"
    effect: deny
    reason: "Block Ruby HTTP server"

Step 3: Block Network Binding Actions

rules:
  - action: network.listen
    effect: deny
    reason: "Block all network port binding"

- action: network.bind
    port: "*"
    effect: deny
    reason: "Block binding to any port"

Step 4: Block Reverse Shell Patterns

Reverse shells are particularly dangerous — they create outbound connections that bypass firewalls:

rules:
  - action: shell.execute
    command_pattern: "nc -l *"
    effect: deny
    reason: "Block netcat listener (potential reverse shell)"

- action: shell.execute
    command_pattern: "nc -e *"
    effect: deny
    reason: "Block netcat with execute flag"

- action: shell.execute
    command_pattern: "ncat *"
    effect: deny
    reason: "Block ncat (nmap netcat)"

- action: shell.execute
    command_pattern: "socat *"
    effect: deny
    reason: "Block socat (bidirectional data transfer)"

- action: shell.execute
    command_pattern: "/bash -i "
    effect: deny
    reason: "Block interactive bash (reverse shell pattern)"

Step 5: Block Tunnel and Proxy Services

An agent might try to expose a local port to the internet using tunnel services:

rules:
  - action: shell.execute
    command_pattern: "ngrok *"
    effect: deny
    reason: "Block ngrok tunnels"

- action: shell.execute
    command_pattern: "cloudflared tunnel *"
    effect: deny
    reason: "Block Cloudflare tunnels"

- action: shell.execute
    command_pattern: "localtunnel *"
    effect: deny
    reason: "Block localtunnel"

- action: shell.execute
    command_pattern: "lt --port *"
    effect: deny
    reason: "Block lt (localtunnel shorthand)"

- action: shell.execute
    command_pattern: "ssh -R *"
    effect: deny
    reason: "Block SSH remote port forwarding"

- action: shell.execute
    command_pattern: "ssh -L *"
    effect: deny
    reason: "Block SSH local port forwarding"

Step 6: Allow Development Servers (Optional)

If your agent needs to start a dev server for testing, create a narrow exception:

rules:
  - action: shell.execute
    command_pattern: "npm run dev"
    effect: allow
    conditions:
      - human_approval: required
    reason: "Allow starting dev server with human approval"

- action: shell.execute
    command_pattern: "npm start"
    effect: allow
    conditions:
      - human_approval: required
    reason: "Allow starting the app with human approval"

# Block all other server starts
  - action: network.listen
    effect: deny
    reason: "All other port binding is blocked"

Step 7: Test and Audit

npx @authensor/safeclaw --simulate

Ask the agent to start a server:

[DENIED] shell.execute: "python -m http.server 8080"
  Rule: "Block Python HTTP server"

Review the hash-chained audit trail:

npx @authensor/safeclaw audit --filter "action:network"

SafeClaw is open-source with 446 tests and works with both Claude and OpenAI providers.

Related Pages

• How to Prevent AI Agents from Sending Webhooks
• How to Stop AI Agents from Running curl or wget Commands
• Deep Dive: Network Policies for AI Agents
• Threat: Recursive Shell Execution