# Ransomware 3AM: 37 machines encrypted and security in doubt

> Bitdefender reconstructs an attack on a British school: 37 machines encrypted, three domains joined and a command pattern that suggests AI assistance.
> By José Maldonado · Cybersecurity · 2026-11-08

Bitdefender published on November 3 its reconstruction of an attack that lasted ten days and ended with 37 machines encrypted at a British school. The detail that has drawn attention is not the damage but the pace: commands that failed were retried every 40 or 100 seconds, with minimal variations between attempts, for hours. For the researchers, that pattern resembles an AI agent loop correcting errors.

The investigation makes clear what it cannot prove. Bitdefender did not have access to the machine from which the attack was launched, so it cannot say which model was used, whether it ran locally or remotely, or even whether a person was copying commands and results between the terminal and a tool. What it documents is the behavior, and the behavior looks like an agent.

The rest of the attack is a familiar playbook. Stolen backup credentials, Mimikatz, two domain administrator accounts created by hand, bidirectional trusts between three Active Directory forests and a Rust encryptor. The [ARTEX campaign with AI-assisted pentesting](https://observatorioblockchain.com/en/cybersecurity/artex-vapor-panda-ia-pentesting-200-organizaciones/) already showed that automation does not need new tools to escalate an incident.

## Ten days inside the network

The report's timeline places the first move on September 10, with scripts that decrypted passwords stored with DPAPI and copied the Chrome credential database. On the same dates, the attacker extracted credentials from three Veeam servers, the system the school used for its backups. That move defines the rest of the story: whoever controls the backups decides how long it takes the victim to recover.

On September 12 it downloaded Mimikatz and renamed it \`sysmon64.exe\`, a name that imitates a legitimate monitoring tool. Two days later it dumped registry hives and created two domain administrator accounts with service names: \`library\` and \`svc\_bkpmon\`. Neither existed in the center's inventory, and neither was detected in time.

Between September 15 and 18 it prepared the exfiltration. Some 40 files from the finance area were packed into Windows registry values and left via HTTP POST from an unmanaged machine serving as command infrastructure. The [Bitdefender analysis of the 3AM ransomware](https://www.bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack/) covers that technical detail and the full sequence.

## Three domains and a bridge to other schools

On September 18 came the most laborious part. With \`dnscmd\` it created DNS conditional forwarders and with \`netdom\` it established bidirectional trusts between Active Directory domains that until then had lived separately. It then ran trust checks with its own scripts, restarted Netlogon and scheduled a PowerShell task that replicated the administrator account across some 40 machines.

That work has a feature that explains why researchers suspect an agent. Joining domains requires configuring both sides correctly; if one detail fails, nothing works, and diagnosis is done by reading the error. The retry pattern every 40 or 100 seconds fits a system that receives the command output and tries another variant, not a person typing.

On September 19 it tried to uninstall Sophos and Bitdefender on the machines reached. Tamper controls blocked it. Even so, it ran \`wbadmin\` to delete system state backups, launched the script loader and deployed the 3AM encryptor. On 37 of the nearly 50 machines reached, the last step worked.

The difference between 37 encrypted machines and 50 reached is no consolation. A modern encryptor settles for the systems that hold data and leaves the rest because the goal is to halt the operation, not destroy the inventory. In an educational center that means records servers, shared drives and administration workstations.

## Why does it look like an agent loop?

The inference rests on three signals. The first is cadence, too regular for a person and too slow for a script that does not think. The second is variation: commands change little between attempts, as if someone adjusted a parameter after each error. The third is the mix of human and automated tasks in the same window of hours.

Bitdefender describes the attack as a hybrid operation and avoids the word autonomy. What it does state is that the hybrid model fits the repetitive work of configuration, where the fix comes when someone reads the error message and decides the next test. The [Dark Reading analysis](https://www.darkreading.com/data-privacy/3am-affiliate-ai-loops-european-school-attack) adds the nuance that matters: there is no proof that AI provided a capability the attacker did not already have.

For the school, the discussion about the model is secondary. The damage came through three known routes: reused credentials with access to the backups, domain trusts that no one reviewed and a protection system that could be disabled if the attacker arrived with sufficient permissions.

## What this case teaches even if AI is not proven

The first lesson is about trust inventory. A school with three separate domains may have a legitimate reason to keep them that way, but each new trust opens a path between organizations that did not exist before. If the team does not have a list of trusts and an alert when one appears, the attacker will create it first.

The second is about backups. Credentials for three Veeam servers allowed the system state to be deleted before encryption, which is the order that turns a recoverable incident into a week of chaos. The [nine-layer guide for SMEs](https://observatorioblockchain.com/en/cybersecurity/ciberseguridad-pymes-nueve-capas-defensa/) insists on the same thing from the other side: a backup that can be deleted with the server's credentials is not a backup, it is a promise.

The third is about separation of privileges. The \`library\` and \`svc\_bkpmon\` accounts worked because one domain administrator can create another. In an educational center with a single team of two people, that separation is difficult, but it is the only way an incident at one school does not become an incident across the network of several.

There remains the question the report leaves open and that will recur in 2027. If a ransomware affiliate can rent a model by the hour to solve its Active Directory configuration errors, the bottleneck of the attack stops being the intruder's skill and becomes the time available. The [PHP2Shell chain in PAN-OS firewalls](https://observatorioblockchain.com/en/cybersecurity/php2shell-fragnesia-pan-os-root-24-cortafuegos/) showed the same trend by another route: custom tools, automatic execution and targets counted in hours.

The practical conclusion does not require resolving the AI question. A school with three domains joined by trusts, uninventoried service accounts and backups accessible from the same server they encrypt is vulnerable with or without an agent. The new part is the speed at which someone can travel that path without quite knowing what they are doing.
