A recent ransomware assault has exposed how modern attackers are fusing human strategy with AI‐like automation to breach UK school networks. In the incident, known ransomware group 3AM–also spelled ThreeAM–employed tools built into Windows, stole backup passwords, and set up fake administrative accounts to infiltrate, maneuver through, and eventually encrypt dozens of machines across a school network. About 37 computers were compromised during the breach.
The targeted infrastructure included 47 systems: approximately 21 servers (eight of them acting as domain controllers) and ten older workstations still running Windows 10. The network originally operated across several unconnected Active Directory domains. Investigators could not trace how the attackers first gained access. Their initial move, observed on September 10–11, involved stealing credentials from three backup servers using a modified public Veam backup script.
Hybrid Attack Model: Manual Oversight Meets AI Repetition
After gaining access, the attackers used Mimikatz—masqueraded as sysmon64.exe—and created two domain admin backdoor accounts under innocuous names like “library” and “svc_bkpmon.” On September 15, they exfiltrated finance files via the Windows registry to a server under their control. One defining moment, on September 13, revealed clear signs of automation: a command repeated after 41 seconds, modified 49 seconds later, then shifted to a different tool after 47 seconds. Retries ranged from 40 to 101 seconds, while task switches unfolded over several minutes. Experts interpret this behavior as a model handling repetitive tasks under human direction.
The attack also displayed errors characteristic of machine memory rather than a live operator’s careful typing. Examples include using incorrect port settings with Veem, inserting bash escape characters into Windows commands, and leaving behind staging files with diagnostic or proof names, plus iterative script versions. Later they merged domains with two‐way trusts using tools like netdom and dnscmd, searched shared folders for other schools’ network info, and attempted to disable security software. Finally, the encryptor—host.exe—was deployed between 05:07 and 05:33 UTC, encrypting at a rate of 12–13 seconds per machine and deleting system‐state backups in the process.
Lessons for Defenders
Security firm Bitdefender calls this a single incident—not yet a trend—but warns such hybrid AI‐assisted attacks may rise sharply in 2027. These attacks tend to generate clear signals: repetitive tasks, numbered or test files, sequential script versions, and frequent retries. Tools leave noise.
In this case, the older toolkit and standard encryptor suggest no cutting‐edge breakthroughs were necessary to pull off the action. However, the consequences were real. Backed up by admins abusing domain trusts, leveraging built‐in tools like netdom, and erasing backups, the attack highlights major defensive gaps.
Key indicators uncovered include backdoor admin accounts (“library” and “svc_bkpmon”), renamed copies of mimikatz masked as sysmon64.exe and svcstat.exe, and various scripts for removing antivirus or escalating privileges. Registry manipulation, batch tasks, resource probes, and a final destructive encryptor round out the profile. Organizations are advised to isolate admin accounts, monitor domain trust changes and netdom usage, safeguard backup servers and maintain separate recovery copies, limit remote registry and ADMIN$ privileges, and deploy endpoint detection across all machines.
What this case shows is that attackers are getting smarter not by inventing new tools, but by repurposing existing ones combined with AI agents or models to reduce manual workload. While people still steer the mission, models appear increasingly in charge of the tedious, repetitive execution. The noise they make—the mistakes, retries, patterns—offers one of the most reliable routes for detection. As schools, hospitals, and smaller targets remain attractive for ransomware actors, investing in observability, attack breadcrumbs, and segmenting trust relationships may be the only way to stay ahead in the evolving hybrid threat landscape. Watch closely: 2027 could usher in an era where AI tools don’t just facilitate attacks—they take on their own signature role.
