A PoE switch is not a switch with a bigger power supply. It is a device that puts 48 volts on a cable before it knows who is on the other end. That fact reshapes every security conversation about it, and almost every hardening guide I read skips it. They tell you to change the default password and put management on its own VLAN, which is correct and also incomplete. Neither stops the port handing 15.4 watts to a device that never authenticated and has no business being in the building.
I found one of those last month. Eight-port switch in a ceiling void above a car park ramp, feeding two cameras and an access point. The customer had a budget problem: 61 W of load on a 60 W budget, and the switch was shedding a port every evening around 19:00. I went up looking for a failing camera and found a PoE-powered Wi-Fi bridge in port 7, drawing 6.4 W, sitting there for what the counter said was 214 days. It was bridging the car park network to a neighbouring building. Nobody had logged into the switch in a year. The port was live the whole time, because a PoE port does not care whether anyone authorised it.
The Port Is Live Before It Asks Who You Are
Detection and classification happen at layer 1. The PSE applies a probe between 2.7 V and 10.1 V across the pairs, measures the current, and looks for a resistance between 19 kΩ and 26.5 kΩ. At 25 kΩ that draws roughly 108 µA, small enough to be safe on an unknown load. If it sees a valid signature it moves to classification, applies about 12.5 V, and reads a second resistor to decide class 0 through 8. Then it applies full voltage. I went through that handshake in an earlier piece, because the sequence is where most PoE surprises come from.
Notice what is missing. There is no authentication step. Link training is not complete, the PHY has barely settled, and the PD is already receiving 44 to 57 V and booting its firmware. 802.1X runs above that, as an EAPOL exchange on a link that is already up, and on the platforms I have worked with it does not gate the PSE. The port authenticates or it does not. The power stays on either way.
I tested this because I did not believe it. A 24-port managed switch, 802.1X enabled globally and on the port, no supplicant, RADIUS returning access-reject. I plugged a PoE tester in and watched it negotiate class 4. 48.1 V open circuit, 15.4 W allocated, port state "unauthorized". It stayed powered the whole twenty minutes. The data path was blocked, which is what 802.1X promises. The power path was not, which is what nobody documents.
For a camera, fine. For a device whose job is to sit on your segment and listen, 15.4 W is a full meal. A small PoE-powered board with a radio draws under 6 W, needs no socket, no UPS, and nobody to open a cabinet. One spare port and a cable is all it takes.
What a Login Actually Buys You
Security advice tends to treat all switch access as equivalent. On a PoE switch it is not, because the highest-privilege action is not reading traffic. It is turning ports off. Someone with admin rights can drop 24 cameras in under four seconds with one command, and from outside it looks exactly like a power failure. That is the difference between an information breach and a physical one.
| Way in | What they can read | Can they cut PoE? | How long before you notice |
|---|---|---|---|
| Open port, no credentials | Whatever the segment broadcasts | No | Never, unless you count ports |
| SNMP v2c read-only | Port table, PoE allocation, MACs, VLANs, firmware | No, but it maps the target | Never |
| SNMP v2c read-write | Everything above, plus the ability to write it | Yes, via the PoE MIB | Not until the next incident review |
| Web or SSH as admin | Config, logs, camera streams if the VLAN is flat | Yes, all ports, in seconds | When someone asks why the gate camera is black |
| Console cable at the cabinet | Everything, including password recovery | Yes, and they can erase the evidence | Possibly never |
| A spare powered port | The segment, passively, for months | No | 214 days, in the case above |
Three of those six rows are things I have found on site, and only one involved a password that had been changed. The spare port gets ignored because it is not a feature you configured wrong. It is a hole you never remembered you had.
The Camera Is the Weakest Thing on the Segment
Here is what makes a PoE network different. The most exposed device on it is also physically tethered to the most privileged one, with a cable you installed, and the tether is what keeps it alive. A camera has no choice about where it connects. Someone with a laptop on the office Wi-Fi has to find the management VLAN. A camera is already inside it, or should be.
Usually it is not. I have lost count of sites where the camera subnet and the switch management address share a broadcast domain. The reason is nearly always the same: someone built a management VLAN, put the switch interface on it, tagged the uplink, then left the access ports untagged in VLAN 1 with the cameras in it. The VLAN exists. It separates nothing. If you are choosing hardware for that job, the difference between a managed PoE switch and a plain one is mostly whether you can enforce this at all.
I got this wrong myself for two years. I used to check the VLAN table, see a management VLAN, and move on. Now I check the trunk's allowed list at both ends, because a trunk that permits all VLANs will carry the management VLAN to a port you did not intend. I also check whether the management address answers an ARP from a camera port. Ninety seconds with a laptop settles it faster than any diagram.
Camera firmware is its own problem and not one I can fix from the switch. But the switch decides how much a compromised camera is worth. If the camera VLAN reaches the management interface, an attacker with a shell on one camera has a hop, and the hop is to the device controlling power to every other camera.
802.1X Doesn't Switch the Power Off
This is the table I wish someone had shown me before I enabled port authentication on a live camera network and spent a Saturday night explaining a recording gap.
| Control | Gates data | Gates power | What it actually buys you |
|---|---|---|---|
| 802.1X with EAP | Yes | No, on every platform I have tested | No traffic until authenticated. Power still flows. |
| MAB (MAC allowlist) | Yes | No | A speed bump. MACs travel in the clear and clone easily. |
| Port security, MAC limit | Yes, on violation | Depends on the action | Stops a rogue mini-switch on one port. Set action to shut down, not drop. |
| Admin-down plus parked VLAN | Yes | Yes, but verify it | The only control that removes power. Check with a tester, not the CLI. |
| LLDP power TLV | No | It negotiates it | Unauthenticated. Whoever is on the far end says what to allocate. |
That last row deserves more than a cell. Fine-grained power negotiation in 802.3at and 802.3bt rides on LLDP, and LLDP has no authentication of any kind. It is a plain layer-2 protocol carrying a type-length-value payload. If the port honours power TLVs, a device on the far end can ask for more than its class implies, up to the class ceiling, and the switch cannot verify the request came from a legitimate PD. I have never seen this exploited. I have also never seen a deployment where it was turned off, mine included.
The failure that bit me was duller. I enabled 802.1X on twenty-two camera ports on a Thursday, using MAB with MACs collected from the ARP table. Eleven passed. Eleven fell to the guest VLAN. The cameras stayed powered, kept recording to their SD cards, and looked healthy on the NVR's device list, because the NVR polled them over a route I had not thought about. What I lost was the recording path, for nine hours. Power was never the issue. That is the point: on a PoE switch, "the device is up" tells you nothing about whether the port is doing its job.
The Order I Do It In
Sequence matters more than the individual settings, because two of these steps reboot PoE and one can lock you out of a switch you are configuring from another city.
- Baseline first. Port by port: link state, PoE class, allocated watts, MAC, budget in use. Without it you cannot tell an improvement from a new fault.
- Count the spare ports twice. Ports with no link, then ports with PoE enabled and no link. Those two numbers are rarely the same.
- Move management off VLAN 1 and confirm it from a camera port. Not from the VLAN table. Unplug a camera, plug in a laptop, try to reach the management address.
- Close the obvious doors. Telnet off, HTTP off where HTTPS is available, SNMPv1 and v2c replaced with v3, and any cloud or remote-access feature you did not deliberately enable switched off.
- Change defaults everywhere, including the camera VLAN's gateway. Vendor defaults are in the manual, which means they are in every scanner's list. Unique per device.
- Disable PoE on unused ports, then measure. Do not assume admin-down removes the volts. I put a tester in the port every time, because I have met a platform where it did not.
- Park unused ports in a VLAN with no gateway and no DHCP. This turns a live jack in a public corridor into a dead one, for free.
- Enable port authentication last, one small batch, with a rollback window. Most cameras do not do 802.1X, so you are relying on MAB, which filters accidents rather than attackers.
- Restrict SNMP and web access to a management subnet, and log every login. A switch that logs admin logins to a syslog server you read beats three features you forgot.
- Re-measure the budget at the end. If you killed four ports carrying phantom load, your headroom improved. Write the new number where the next person will find it.
Step eight is where people get hurt. Enabling 802.1X on a PoE switch is not a change you undo with a rollback timer, because the switch restarts its PoE controller and every PD power-cycles. On the hardware I test, a full restart runs about 52 seconds of boot, plus staggered port bring-up, plus 45 seconds for a camera to come back and stream. Call it 110 seconds of nothing, multiplied by however long it takes you to notice. If you are doing it remotely, do it at 14:00 on a Tuesday. The same restart is what makes dual power supplies less of a safety net than the datasheet suggests.
Questions I Get Asked
Does an unmanaged PoE switch need hardening?
A different kind, because it has no login to protect. What it has is a live port on a wall. Most unmanaged units cannot disable PoE per port, so your controls are physical: locked cabinets, cable management that hides spare jacks, and not leaving a 24-port switch in a corridor with eleven unused ports. I would rather have a managed PoE switch with the spares switched off than one where I cannot turn anything off.
Is a separate management VLAN enough on its own?
No, and it is the most common false confidence I run into. A management VLAN tagged on a trunk that carries all VLANs, with access ports still untagged in VLAN 1, separates nothing. The test is not the VLAN table. The test is whether a device in a camera port can reach the management address.
Can 802.1X stop someone plugging a laptop into a spare port?
It can stop the laptop passing traffic, and that is worth having. It will not stop the port powering whatever is plugged in, so a PoE-powered listening device still gets fed. If a port matters enough to authenticate, it matters enough to have power switched off when idle.
How much power does a rogue device actually need?
Less than you would guess. Class 0 is 15.4 W at the PSE, 12.95 W at the device, which fits a small computer with a radio. Even the 6.4 W bridge in that car park ran for months. The port does not audit what it feeds. It audits whether the resistance looks like a PD.
Should I worry about the LLDP power negotiation?
Worried enough to know whether it is enabled. Not enough to lose sleep. The attack needs someone already on the segment who can craft frames, and at that point you have worse problems. But if your platform lets you disable extended power negotiation on ports that never need it, that is free surface reduction.
Where I Would Spend the Next Hour
If you take one thing from this, take the ordering. The instinct is to start with passwords, because every checklist says so first, and passwords are the least interesting control on a PoE switch. The interesting one is the port that is powered, reachable, and forgotten. Find those, count them, kill the power, write the number down. Above that car park ramp, the number was one. It had been one for seven months.
I would also stop treating the power budget as a capacity figure and start treating it as an audit trail. A budget that is 2 W short is telling you something. This time it told me there was a device nobody had installed on purpose, and the finding arrived disguised as a load-shedding complaint. That is how these things usually arrive: not as an intrusion, but as a device that stopped working for reasons nobody can explain.
If you have a switch where the port count and the PoE allocation do not add up, or you have measured something that contradicts what is above, I would like to hear it. The 802.1X behaviour in particular varies by platform. My sample is not big enough to be a standard, and I would rather be corrected than quoted.