The Ransomware Loop Just Hired a Co-Pilot
A UK school estate was taken down in September by the 3AM ransomware crew. That part is familiar: domain-admin theft, lateral movement, backups deleted, 37 of 47 systems encrypted. The new part is in the forensic timing.
On Nov 3, Bitdefender published a case study of the intrusion. BleepingComputer picked it up the same day; so did Dark Reading and Hackread. Bitdefender does not say a language model ran the intrusion. It says the retry cadence looked like something that was not a person.
What the shell log showed
The clearest signal is on Sept 13. From Bitdefender's reconstruction:
- The attacker reran a command unchanged 41 seconds later.
- Then altered one call 49 seconds after that.
- Then switched to a different tool 47 seconds later.
- Across the rest of the operation, failed commands were retried every 40 to 101 seconds.
A human operator has moments of thought, pauses to read, moments of frustration. This did not. It also made the kind of mistake a person would not make and an autocomplete would: bash-style escape characters in Windows commands on three separate days, a wrong port and wrong login method against a Veeam server, numbered iterations of the same script (av_remove, av_remove2, av_remove3). Test files left on disk were named for AI developer idiom — psexec-ok, da_proof.txt.
Bitdefender is careful about what this shows. It writes that it "could not confirm whether a model ran the loop or an operator relayed commands between a terminal and an AI tool." No specific model or product is named. BleepingComputer calls the intrusion "not a confirmed autonomous ransomware attack."
That is the right level of certainty. What is new is that the question is now a forensic one.
The rest of the attack was ordinary
Everything else in the Bitdefender writeup is standard ransomware tradecraft, and that is the point:
- Sept 10-11: Backup server credentials stolen from three Veeam backup servers using a modified public script.
- Early September: Mimikatz downloaded and renamed to
sysmon64.exe. Two domain admin accounts created with names chosen to blend in —libraryandsvc_bkpmon. - Sept 15: About 40 finance files staged in Windows registry values, then exfiltrated to an attacker-controlled server.
- Sept 18: Three separate Active Directory domains joined with two-way trusts using
netdomanddnscmd. Shared folders searched for other schools' names and networks. - Sept 19: Attempted to remove Sophos and Bitdefender endpoint protection. Anti-tamper blocked it. The Rust-based
host.exe3AM encryptor ran between 05:07 and 05:33 UTC, taking 12 to 13 seconds per machine.wbadmindeleted system-state backups on the way out.
None of these steps required a new primitive. Living-off-the-land with PowerShell, WMI, netsh, reg, and rundll32, plus Mimikatz and Impacket-style remote execution. 3AM itself is a Rust ransomware sold as a service — the group's affiliates conduct the intrusions and split the proceeds. The operator who hit the schools is not named.
Why the retry cadence is the story
The reason Bitdefender's writeup matters is not that an AI did ransomware. It is that the telltale is now in the shell log, not just the binary.
For the last two years, the AI-in-ransomware question has been argued at two distance extremes. On one end, attacker-side exposed workspaces — the SOCRadar / n0n OpenCode leak counted a $7,216 model spend across 154 model entries at seven providers. On the other end, vendor attributions — CrowdStrike's Vapor Panda tracking of ARTEX and CyberStrikeAI from last week. One shows the attacker's rig; the other shows the attacker's brand.
The 3AM schools case is the middle layer. It is a defender's post-incident reconstruction of a real intrusion, in a real environment, where the AI signal is readable from how the shell was driven. A 41-second-then-49-second-then-47-second decision cadence is a per-second observable. A Bash-syntax leak into a Windows command is a per-command observable. An iterated av_remove script series is a per-filename observable. Each of those is a forensic artifact a backup server or EDR can preserve.
What it means for the next one
If 3AM's affiliate is not the only one running a loop like this, and the economic logic of that argues against — hybrid loops cut dwell time and operator cost at once, and the early signs are in the writeup — then the forensic question for the next intrusion is not whether to look for AI assistance. It is where the retry cadence shows up in your logs.
Backups are still the operational pivot: Bitdefender notes Sophos and Bitdefender anti-tamper blocked the AV-removal attempt, and that holding the backup credential boundary and the AD trust boundary is what buys defenders the time they need. Bitdefender expects hybrid attacks to grow through 2027.
The schools' Sept 19 encryption ran for 26 minutes end-to-end. The retry loop that preceded it ran for nine days.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.