All articles

NewsNovember 3, 20265 min read

3AM ransomware hits UK schools in hybrid human-AI attack

Bitdefender: a 3AM affiliate encrypted 37 machines at a UK school estate, with commands that look generated by an AI loop. Attack chain and what to do now.

Illustration about ransomware encrypting data, with the words Encrypting Ransomware, a businessman and icons of files, padlocks and money

An affiliate of the 3AM ransomware compromised the network of a school estate in the United Kingdom between 10 and 19 September 2026: from an unmanaged host it reached about 50 machines and encrypted 37 of them. Bitdefender reconstructed the incident in a report published on 3 November, covered by BleepingComputer, Dark Reading and Infosecurity Magazine (Dark Reading describes it as a “European” school, the other two outlets as a UK one).

What is new is the way the attacker worked: according to the researchers, some commands were written and revised by a language model running in an agentic loop, while a person chose the next steps. Bitdefender calls it the first case of this kind it has documented in Europe, but is clear about the limits of its analysis: it had no access to the attacker’s machine, so it cannot say which model was used, whether it ran locally or remotely, or even whether a person was copying instructions and results by hand between the implant and the model. As Dark Reading points out, the researchers do not claim the attack was fully automated.

Why AI is suspected: the rhythm of the commands

The clue is in the timing. On 12 and 13 September the attacker tried to reach other hosts from a machine where SQL Server was the parent process, sending commands to the WMI service with stolen administrator credentials. According to Infosecurity Magazine, a PowerShell command was rerun unchanged after 41 seconds, changed in a single place 49 seconds later and, another 47 seconds later, rewritten from scratch with wmic, the older command-line interface to WMI.

The next nine retries followed the same pattern: 40 to 101 seconds between attempts, whatever the size of the change, while moving on to a new task took minutes. For Bitdefender, this split, long gaps between tasks and short ones between retries, is what you would expect when a person decides the next step and a model handles the retries. Earlier, simpler commands arrived 2 to 16 minutes apart, at the pace of someone reading the output and thinking.

The attack chain

Bitdefender could not determine the initial access. The phases reconstructed, combining the sources:

  • 10 September, backup credentials: using a modified public PowerShell tool, the attacker extracted and decrypted (through the Windows Data Protection API) the credentials stored on three Veeam backup servers; they also copied a Chrome database containing credentials and used Mimikatz and LSASS memory-dumping tools;
  • reconnaissance: mapping of domain relationships and searching logon-script folders for references to other schools;
  • 14 September, persistence: two domain administrator accounts created by separate routes, named library and svc_bkpmon according to BleepingComputer;
  • exfiltration: about 40 finance-related files copied with robocopy and sent to the command-and-control server via HTTP POST requests. According to Infosecurity Magazine, Bitdefender describes the use of the Windows registry as a transport channel as “a strong indicator to hunt for”; the sources give no further details;
  • 19 September, domain trusts: three previously separate Active Directory domains were joined with two-way trusts and the backdoor account’s access was pushed across the whole estate in about three hours, with diagnostic probes and a restart of the service that maintains the secure channel between domains. For the researchers this is the clearest example of the hybrid approach: mechanical, repetitive work where the fix comes from a person reading the error;
  • impact: attempts to remove endpoint protection (Sophos and Bitdefender, blocked according to BleepingComputer), deletion of system-state backups and launch of 3AM’s Rust-based encryptor, which succeeded on 37 machines.

Matching MITRE ATT&CK techniques, as an indicative mapping: T1047 (WMI), T1059.001 (PowerShell), T1003.001 (LSASS memory), T1555.003 (credentials from web browsers), T1136.002 (domain account), T1484.002 (domain trust modification), T1562.001 (disable or modify tools), T1041 (exfiltration over C2 channel), T1490 (inhibit system recovery), T1486 (data encrypted for impact).

3AM is a ransomware-as-a-service operation observed since 2023; according to BleepingComputer it has been linked to the Conti and Royal ecosystems, and its affiliates have also used email bombing and fake IT support calls to get into networks.

Impact for European and Italian companies

The techniques are not new: theft of backup credentials, rogue domain administrators, domain trusts, legitimate Windows tools. What changes is the trail left behind: close, regular repetitions of near-identical commands, which a model produces effortlessly. Bitdefender believes the hybrid approach makes economic sense for attackers and expects it to become more common in 2027; on the same day the UK NCSC urged critical infrastructure to prepare for AI-powered ransomware by 2028.

The case is directly relevant to organisations with several sites and several domains (school groups, companies with branches, healthcare providers) and to anyone running backup servers with stored privileged credentials. For entities in scope of NIS2, an incident like this triggers the required notification obligations.

What to do now

  1. Treat backup infrastructure as a target: backup servers separated from the production domain, dedicated credentials, immutable or offline copies, alerts on access to stored credentials.
  2. Keep each domain’s administrative accounts separate and never store privileged credentials for several domains in one place; no passwords saved in browsers on servers.
  3. Watch Active Directory: alerts on trust creation (event 4706), new accounts and additions to privileged groups (4720, 4728); look for accounts such as library and svc_bkpmon.
  4. Restrict remote execution: WMI, PowerShell remoting and administrative shares only where needed; block interpreters and shells on machines that never use them, and flag SQL Server launching PowerShell or wmic.
  5. A protection agent on every machine, with tamper protection enabled: this attack started from an unmanaged host.
  6. Hunt for machine rhythm: sequences of near-identical commands at regular 40–100 second intervals, robocopy copies of administrative folders, HTTP POST to unknown destinations.

How ARXDOME helps

ARX Watch finds devices without a probe on the network, like the unmanaged host this attack started from. ARX Sentry recognises critical servers, such as backup servers and domain controllers, and brings account changes, explicit credential use and suspicious executions into the SOC. ARX Shield shows the protection status of every machine, installs antivirus where it is missing and isolates a suspicious machine; ARX Trace reconstructs the chain with MITRE ATT&CK mapping.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles