How to Detect VPS Abuse From the Hypervisor Without an Agent
Find outbound DDoS, spam, scans and noisy neighbours from the node, with no agent inside customer VPS's. What you can see, and the commands.
By Frabs · 1 November 2026 · 5 min read
The short answer: you can see most VPS abuse from the node itself, because every packet and every CPU cycle a guest uses passes through the host. Watch each guest's network interface and its resource use from the hypervisor, and you can spot outbound floods, spam, scans and noisy neighbours without installing anything inside a customer's machine.
You often have no choice in this anyway. Customers own their VPS's, you may not have access to them, and the machine that is compromised is the last place you should trust for reporting on itself.
Why Monitor From the Host, Not Inside the VPS?
An agent inside each guest has real drawbacks for a hosting provider:
- You need access you don't have. Unmanaged customers run what they like and can remove anything you install.
- It cannot be trusted once the guest is compromised. An attacker with root can stop or fool it.
- It is one more thing to support, across every operating system your customers choose.
- Customers may not want it, and they are entitled to their privacy.
The host has none of these problems. It sits beneath the guest, so the guest cannot hide its traffic from it.
What You Can See From the Node
Network traffic for each guest. Every VPS has a virtual interface on the node, such as tap101i0 on Proxmox VE or vnet0 with libvirt. Whatever it sends passes through there. You can see packet and byte rates, protocols, destination ports and how many hosts it is talking to.
Resource use for each guest. Each VPS runs as a process in its own cgroup, so CPU time, memory and disk activity can be read without entering the guest.
The health of the node. CPU, memory, disk and interface errors show whether one guest's behaviour is actually hurting the others.
What you cannot see: the contents of encrypted traffic, what is running inside the guest, or customer files. For abuse detection you rarely need them. The pattern of the traffic usually says enough, and staying out of the guest is the point.
How to Spot Each Type of Abuse
To find the interface for a guest on Proxmox VE, where VM 101's first interface is tap101i0:
qm config 101 | grep netOr with libvirt:
virsh domiflist <guest>To watch one guest's traffic:
tcpdump -ni tap101i0 -c 200To get its counters with libvirt:
virsh domifstat <guest> vnet0Outbound DDoS
A very high packet rate towards one or a few destinations, usually UDP or SYN packets, with few replies coming back. This is the type most likely to hurt everyone quickly, because it can fill your uplink and get addresses null-routed.
Spam
Many connections to port 25 on many different hosts. A VPS that does not normally send mail and suddenly does is a strong sign. You can count it without reading any mail.
Port Scanning
Connections to one port across a large number of hosts, or to many ports on one host, with most going unanswered. Many hosts on random high ports with answers looks more like peer-to-peer traffic, so check before you act.
Brute Force
Repeated new connections to login ports such as 22 or 3389 across many targets, in short sessions that end quickly.
Noisy Neighbours
A VPS holding its CPUs or disk flat out for hours can slow everything else on the node. Read the CPU steal time inside a neighbouring guest if you have one, or look at the node. For CPU:
top -d 5For disk:
iostat -x 5The important question is whether anyone is actually affected. A VPS using its full allocation while the node has room to spare is using what it paid for. Limit it only when others are slowing down.
The Problems You Run Into
Knowing which customer it is. Address and interface mappings change as guests move, are rebuilt or recycled. You need to know who held that address at that time, not who holds it today.
Thresholds that cry wolf. One number for every VPS catches game servers, VPNs and backups. An alert nobody trusts soon gets ignored.
Telling victims from attackers. A VPS under attack sends replies to forged addresses, which can look like an attack of its own. Punishing it makes things worse.
Scale. Watching one guest by hand is easy. Hundreds, all day, with history, is not something a person can do, and home-grown scripts tend to be replaced with something more dependable sooner or later.
Leaving your firewall alone. Anything you add has to work alongside the tools you already run, and must not take the node down if it fails.
How Frabs Does This
Frabs Hypervisor is built for exactly this. One sensor runs on each node, on KVM, containers and Xen, and nothing is installed inside a VPS. It reads the node's setup and proposes a plan, and nothing is enforced until you have confirmed it.
It watches each VPS's traffic and resources, names the VPS and the customer, and limits abuse by the rules you set. If the sensor stops, by default the network changes it made are removed and traffic carries on, so a problem with Frabs does not become an outage for your customers. You can try it free for 30 days on one node.
Frequently Asked Questions
Can you monitor a VPS without installing an agent?
Yes. Every VPS's traffic and resource use passes through the node it runs on, so you can see network activity, CPU, memory and disk from the host without touching the guest.
How do I find which VPS is sending spam or a DDoS?
Find the guest's virtual interface on the node, such as tap101i0 on Proxmox VE, and capture its traffic with tcpdump. Sending to port 25 on many hosts suggests spam, and a very high packet rate to a few targets suggests a flood.
What is a noisy neighbour on a VPS?
A VPS that uses so much CPU or disk that other VPS's on the same node slow down. Short bursts are normal, so a limit should only follow use that lasts for hours, and lift as soon as it calms down.
Can the hypervisor read what is inside a customer's VPS?
Abuse detection does not need to. Watching traffic patterns and resource use from the host is enough, and it can be done without opening customer files, email or databases.