The ugliest part of the SCKit campaign Microsoft documented on November 13 isn’t the backdoor. It’s that nobody did anything wrong according to the tooling. A technician’s session ran ScreenConnect, which it does every day. A ScreenConnect client service called home over 443, which it does every day. A command shell spawned under that service, and that is where the day stopped being normal, except by then the attacker already had a hands-on-keyboard operator sitting inside your management plane.
No CVE here. Nothing to patch. ScreenConnect (ConnectWise Control) behaved to spec. SCKit just arranged to be the one holding the console. So I spent a week running two defensive postures side by side against the behaviour Microsoft described, scoring each round on what actually matters when an intruder is wearing your help desk’s clothes.
The two contenders
Corner A, Detect: instrument everything. Defender XDR hunting queries, EDR, a tight allowlist of known-good ScreenConnect instance IDs, alerts on new service installs. You keep the flat, convenient support model and you watch it like a hawk.
Corner B, Isolate: treat remote-support tooling as its own blast-radius domain. Technicians reach production through a segmented support plane; the ScreenConnect relay and clients can’t freely route to domain controllers, backup servers, or each other. You trade convenience for containment.
Running score below. First, the attack both have to survive.
What SCKit actually does
Per Microsoft’s analysis, initial access leans on social engineering of the people running support, not the end users receiving it, phishing lures and trojanized installers that stand up an attacker-controlled ScreenConnect instance, plus abuse of exposed or loosely-governed existing instances. Once a client agent is present, the operator has what they came for: an interactive session under a technician’s context, a trusted outbound channel to a relay that looks like legitimate SaaS traffic, and a service that restarts itself after reboot. That’s your persistence, your C2, and your privilege, delivered by a signed binary your allowlist probably trusts.
Then it’s the usual: credential theft from the technician’s session, lateral movement into the management tools that session can already touch, and, because this is a technician, a network position most of your Conditional Access and PAM controls assume is safe. That assumption is the whole exploit.
A compromised end user is a door into one room. A compromised technician session is a door into the room where the keys are kept. Stop treating them as the same incident.
Round 1, Resisting initial access
Detect does almost nothing here, and it’s honest to admit it. A silent installer dropping a legitimate signed agent trips no exploit signature. You might catch the phishing email; you might not.
Isolate doesn’t stop the phish either, but a rogue agent that lands in a segmented support subnet reaches far less. The attacker gets a foothold, not a kingdom. Isolate 1, Detect 0.
Round 2, Spotting persistence
This is Detect’s best round, and it earns it. The tell is a ScreenConnect client service spawning a shell or an unexpected child process. If you have Defender XDR:
No XDR? The logic is vendor-neutral. Alert on: creation of a Windows service whose binary path contains ScreenConnect Client, any process tree where that service is the parent of an interpreter, and outbound connections from the client to a relay host that isn’t your sanctioned instance. Sysmon Event ID 1 (process create) filtered on the client service as parent, plus Event ID 11/13 for the install footprint, gets you most of the way.
Detect finds this reliably if you wrote the query before the incident. Detect 1, Isolate 1, but read the fine print: it finds it, it doesn’t stop it. You’re alerting on an operator who is already inside.
Round 3, Blast radius once they’re in
Here’s where the week got grim. In the flat model, a single rogue agent on a technician box could reach the same management surfaces the technician could, which is to say, nearly everything. Detection gives you a timeline of the destruction. Isolation gives you a wall the destruction runs into. One of those is a defence. Isolate 2, Detect 1.
Round 4, Finding what’s already there
Both postures need you to inventory live agents and separate yours from theirs. Instance ID is the discriminator, your tenant’s clients share a known relay and thumbprint; SCKit’s won’t.
Cross-reference every discovered instance ID against your one sanctioned value. Anything else, unknown relay FQDN, a service name with an instance GUID you don’t recognise, an install under a user profile path, is guilty until proven otherwise. Validate, don’t just kill: capture the relay host and binary first, then remove the service and the folder. This round is a draw; both need the inventory.
What to keep logging either way
- Service creation events where the binary path mentions ScreenConnect, this is your single highest-value signal.
- Outbound 443/8041 from client hosts to any relay that isn’t your instance’s FQDN.
- Session start anomalies: connections outside support hours, from new geographies, or to machines a technician never touches.
- Process ancestry: the client service as parent of any shell or scripting host.
The verdict
Isolate takes it, 2 rounds to 1. Detection is necessary and I’d never run without it, but a hunting query is a smoke alarm, not a fire door. It tells you the technician plane is burning; it doesn’t keep the fire out of the server room. The architectural answer to the brief’s question is yes: remote-support tooling belongs in its own segment, reaching production through brokered, logged, time-boxed paths, not sitting flat alongside the crown jewels because it’s convenient for the help desk.
The one case I’d pick Detect alone: a small estate where network segmentation is months of change-control you don’t have, and a well-written set of queries plus instance-ID allowlisting is deployable this afternoon. Ship that today. Then start building the wall, because SCKit proved the thing you trusted to do the fixing is also the thing they’ll use to do the breaking.
