Bitdefender suspects AI helped the hackers who attacked UK schools

Between two commands, 40 to 101 seconds went by. It was the same attempt, relaunched again and again, identical or slightly modified, on the network of a group of schools in the United Kingdom. That steady rhythm tipped off Bitdefender's analysts. On November 3, the Romanian security vendor published its investigation into the attack, carried out by an affiliate of the 3AM ransomware, meaning a hacker who rents this encryption software from its authors.

The intrusion ran from September 10 to 19, 2026. Of the roughly 50 machines reached, about 37 were encrypted on the morning of September 19. Financial data was stolen. How the attackers got in is still unknown.

Nine days to take control of the network

From the first days, the attackers went after the backups. Using a modified public script, they extracted the credentials stored on three Veeam backup servers. They also copied the passwords saved in Chrome, then used Mimikatz, a classic credential-theft tool, renamed to go unnoticed. They then created two administrator accounts with names that blend into a school environment: “library” and “svc_bkpmon”.

The network was originally split into several separate domains. On September 13, they searched for references to other schools. On the 18th, they created trust relationships between three of these domains to widen their access. On the 19th, they tried to uninstall the Sophos and Bitdefender protections. The anti-uninstall safeguards blocked the attempt. Encryption took place anyway, and a system backup was erased.

What gives away an AI?

Bitdefender points to several clues. Writing habits specific to Linux mistakenly showed up in Windows commands, on three different days. The call to the Veeam software used a port and an authentication method that do not match the official documentation. These are typical mistakes of knowledge recited from memory. Finally, the repeated attempts followed one another quickly, while task changes took several minutes.

According to the vendor, these clues strongly support the hypothesis that a language model wrote the commands. There is no proof, however, that an AI agent chose each action on its own: a human copying and pasting between a console and a chat window could produce the same rhythm. Bitdefender describes a hybrid operation, in which a human sets the objectives and the AI carries out the repetitive tasks. It presents this as the hypothesis defenders should work from, not as an established fact. The vendor can say nothing about the model used: the attackers' machine was outside its visibility.

Good habits, AI or not

Nothing is new in the techniques, Bitdefender insists: only the traces look different. Like other ransomware, such as SafePay, this one targeted backups first. The vendor recommends:

• separating the administration accounts of each domain and requiring approval for any new trust relationship;
• isolating backup servers and keeping copies that no domain account can delete;
• restricting remote execution (WMI, remote registry, scheduled tasks) to known administration workstations;
• installing a security agent on every machine, without exception;
• monitoring for credential reads in backup databases and for the creation of trust relationships between domains.

The United Kingdom is already preparing: its national cybersecurity center, the NCSC, has urged critical infrastructure operators to get ready for AI-driven ransomware by 2028. This case suggests that some hackers have already started experimenting.