In August a client called me about windows that kept opening on their laptop by themselves. Underneath them was a remote-access trojan that had been there for six months. I wrote that up in I got called about a popup. That post covers what I found on the disk.
This one covers what the malware does when it runs. In October I went back for the actual samples, sealed them in a lab, and ran them on purpose to watch. That’s called dynamic analysis, or detonation. Claude Mythos 5.1 did most of the analysis.
Identifying details are withheld. The technical content is exactly as observed.
Why run malware on purpose
I’d already spent an evening on static analysis: reading the files without running them. That tells you what a program can do, not what it does, in what order, how fast, and what quietly fails on a real Windows machine.
The lab corrected that reading more than once. A persistence mechanism an earlier review had called optional switched on by default. The implant came back from the dead faster than Windows’ restart settings alone would explain. And it left a backup copy of itself in a folder nobody had predicted.
Reading the code tells you what’s possible. Running it tells you what’s true.
The lab
- A virtual machine with no way out. UTM on an M4 Pro Mac, running Windows 11 on ARM, with no network card at all (not disabled, absent), no shared clipboard and no shared folders.
- It calls home to itself. The attackers’ server names were pointed back at the VM’s own address, so every attempt to reach them landed on a listener I controlled.
- Everything recorded. Sysmon and Process Monitor logged every process, file, registry and network event, against an Autoruns baseline.
- One way in, one way out. Samples went in on read-only virtual CDs. Results came out by cloning the VM’s disk after shutdown and reading the clone offline on the Mac, without ever booting it.
- A clean slate every time. Each run started from the same clean snapshot.
Five runs, between 12:38 and 2:05 a.m. on 15 October. For scale, run two lasted about eight minutes and Process Monitor logged about 2 million events in it.
Before anything ran, a separate review checked every sample for code that detects virtual machines, debuggers or monitoring tools, or tries to break out. There was none, so what follows is the malware behaving normally, not putting on a show for a sandbox.
Who did what
The static analysis that evening ran on Claude Opus 5.5 in Claude Code. Opus also built the lab, staged the first run and started it.
Eight minutes into run one, the session moved over to Claude Mythos 5.1, which I’d just been given access to. From there Mythos ran the lab:
- It checked the keylogger timing test and ran the tests where we tried to kill the implant.
- It wrote the scripts for runs two through five, and pulled results off the cloned disks when the results drive failed.
- It decompiled the popup from its .NET bytecode when the usual decompiler wouldn’t run.
- It kept up the timestamped action log, lined up with my screen recording, and wrote the report.
The raw logs, nearly 2.9 million Process Monitor events for run one alone, went to Opus subagents so Mythos could keep the lab moving.
My part was deciding what to run and when to stop, recording the screen, and fighting the popup by hand the way my client had.
Run five finished and its disk is preserved, but it hasn’t been analysed yet, so nothing below comes from it.
The model did the analysis and most of the driving. What to run, and when to stop, were my calls.
It locks administrators out
Installing it created two Windows services, “Windows Sync Service” and “Windows Sync Monitor,” running as SYSTEM, which is the account that sits above administrator. In the same instant, before it had contacted any server, it rewrote the services’ permissions. Administrators can see them, but can’t stop, reconfigure or delete them. An elevated PowerShell got “Cannot open WinSyncAgent service.” That was the “Access is denied” from August, reproduced on demand.
Then we tried to kill it:
- Killing the keylogger: back in 7.6 seconds.
- Force-killing the main process (that works, though stopping the service doesn’t): the monitor service restarted it 2.0 seconds later. Windows’ own recovery setting would have restarted it at five seconds anyway.
One install produced 11 of the 12 persistence mechanisms in the code, including two services, a boot task, an hourly task, an hourly WMI subscription, a backup copy of itself and an encrypted copy of its settings. The tasks and the WMI subscription all run a “heal” command that puts the rest back.
Microsoft Defender was running the whole time, with real-time protection and tamper protection on. It scanned the files as they were written. It flagged nothing, not one component.
The keylogger runs on a 60-second clock
It records every keystroke in every application, plus anything copied to the clipboard. For each keystroke, it asks Windows’ accessibility interface whether the cursor is in a password box, and marks the ones that are. It has a list of programs to ignore. On the client’s machine that list was empty.
Keystrokes are bundled into batches, encrypted so that only the operator can read them, and written to disk about 60 seconds after the first keystroke. In the lab I typed a test string into Notepad and the batch appeared 60.05 seconds later. A batch stays on disk until the attacker’s server confirms it arrived, and then it’s deleted.
That clock settled a question from August. A few minutes into my response, I typed a message containing a password. Did they get it?
On the real laptop, the last batch was written under two minutes after I started typing, which fits my first few messages. The password went in about three minutes later. Had it been recorded, a batch would have hit the disk within about a minute, and I didn’t stop the keylogger for more than four. No further batch was ever written. The password was not captured, and the lab’s measured 60 seconds is what lets me say that with confidence rather than hope. Changing it anyway is still good hygiene.
The uncomfortable part: the attacker’s connection was live for the first ten minutes of my session, until containment cut it. That first batch, holding my opening messages, probably went straight to them.
The popup has exactly one exit
Run two launched the popup the way it ran on the laptop: a scheduled task at sign-in, as the ordinary user.
Within seconds a dialog appeared, styled as a Windows system message. The title read “Phone Link could not finish synchronizing.” The body said “this session cannot continue until Phone Link setup is completed,” and there was a single button: “Complete Synchronization.” Behind it, it opened the real Phone Link app and pinned it on top.
Underneath, it works like this:
- Every four seconds it checks whether a phone is paired. It looks in Phone Link’s data folder and in Windows’ Cross Device folders and registry. That came to 121 checks in eight minutes, and every one said no.
- If a phone is paired, it cleans up after itself. It removes its startup entry, puts Task Manager back, and closes. Nothing visible is left.
- If not, it digs in. It adds itself to startup, tries to switch Task Manager off, swallows Ctrl+Shift+Esc, and relaunches Phone Link every time you close it.
I closed Phone Link nine times. It came back nine times, each within six seconds. Afterwards I told Claude, “Exactly how I remember it.”
One lab detail: the Task Manager switch is a registry policy, and the write was denied because the program runs without admin rights. Catching Ctrl+Shift+Esc did the job anyway.
It also has no network code at all. It steals nothing and sends nothing. It exists to make you pair your phone.
Every way off that screen runs through pairing your phone.
That’s because Phone Link puts your phone’s text messages on the PC, and the attackers already controlled the PC. Pair the phone, and every two-factor code sent by text lands on a screen they can read. That’s what Phone Link was there for: getting around SMS two-factor by intercepting the codes. On this laptop it never got that far, because the phone was never paired.
What it was built to do
This is custom code, not a repackaged commercial tool. It’s written in Rust, and the build paths left inside it show a Windows user named BTC and a project folder A:\Project\RMM2. RMM means remote-management software, the kind IT companies use on their clients’ machines. This is somebody’s own, version two.
- 30 commands. Screenshots, file transfer, a remote shell, running and killing programs, services, power, inventory, remote desktop, keylogger control, self-update, and a command to erase itself.
- It talks over a secure WebSocket, which is an HTTPS connection that stays open. To a firewall it looks like ordinary web traffic.
- A hidden administrator account,
WinSyncSvc, kept off the sign-in screen. The operator can order a new random password at any time and receive it encrypted. That’s a standing admin login that works over remote desktop or file sharing whether or not the implant is still installed. - A private channel waiting for passwords from a separate component, most likely one that sits in the Windows sign-in process. Nothing connected to it in the lab, and I never found that component.
Two more details for defenders. Given a fake server with a certificate the VM trusted, it connected without complaint, and its opening request carries no credentials and no browser identifier, a pattern network tools can match. And its backup way of finding servers does its own DNS lookups, which Windows’ DNS logging never sees.
The people behind it
The attackers’ TacticalRMM server domain was registered on 14 May 2026 at 15:14 UTC. This laptop enrolled about two hours later as “Client 1, Site 1”, the first client on a freshly built server. That fits the Transfer old bots 13may label in the previous post.
The implant’s own server is older. Public sandbox and crawler records show dozens of other agents checking in to it between mid-May and the end of July, using an older version of the protocol. On 15 October all of it was still live, and the implant server’s certificates had been renewed within two days of my cutting the laptop off in August. I’m not publishing the domains or addresses here.
The samples also carried tokens for the attackers’ own servers, sitting in configuration files and installers. They would have let me into their servers. I didn’t use them.
They’d been knocking since November
Going back through the mailbox turned up 13 lure emails in five waves: late November 2025, early December, 21 January 2026, late March and late April. Every one was a genuine Google Drive share notification, under a famous law firm’s name spelled with lookalike letters, and every one was about a debt: “Settlement Letter Arrived — Take Steps!”
The January wave arrived on the same day as the earliest trace of the attackers on the laptop. It’s most likely the one that worked. The March and April waves arrived after the laptop was already theirs.
What the lab changed about cleanup
- Disabled isn’t deleted. The hidden account is managed by the implant, and it can get a new password on command. It has to be deleted, along with the setting that hides it from the sign-in screen.
- An administrator can’t remove this. The services have to be stopped from a SYSTEM-level shell or offline, in order: stop them, strip their restart settings, remove the WMI subscription and tasks, then delete the services. Out of order, it puts itself back.
- Find the spare. The backup copy of itself in a second folder has to go too.
- Rebuild if you can. After months of something running as SYSTEM with locked permissions, a clean rebuild is the only end state that doesn’t depend on finding every last piece.
What I’d tell a business owner
- Run it, don’t just read it. Static analysis gave me a map. The lab showed where it was wrong, and answered the client’s real question, whether the password got out, with a measurement instead of a guess.
- Antivirus silence isn’t evidence. Defender with every protection switched on watched this install itself, lock out administrators and start logging keystrokes, and it said nothing.
- In a live response, assume they’re watching. The attacker was connected for the first ten minutes of mine. Cut the machine’s connection first, and type nothing sensitive until you have.
- A popup with one way out is the attack. Pair your phone, call this number, install this update. If the only way to make it stop is the thing it’s asking for, don’t do it.
- Text-message two-factor is only as safe as every screen your texts reach. For anything with money behind it, use an authenticator app or a passkey.
- AI made this a one-night job for a one-person shop. It’s still a tool, and someone has to check its work: its earlier readings got corrected several times that night.
If a symptom on one of your machines looks manufactured, it’s worth an hour of somebody watching it properly before anyone clicks the button it wants clicked.