Bitdefender reports that a 3AM ransomware affiliate encrypted about 37 machines in a UK school estate after stealing credentials and expanding administrative access. Marius Baciu’s November 3 investigation traces activity from September 10 through 19. The attacker reached roughly 50 machines from an unmanaged internal host, outside the investigators’ endpoint visibility.

Evidence status: a real intrusion reported by the investigating security vendor. The report includes sanitized command logs and identifies the 3AM encryptor. Its suggestion that an AI agent handled some repetitive work is an interpretation of that behavior. The initial access route remains unknown.

Backup access became a route to wider control

The first recorded September activity targeted three Veeam backup servers. Bitdefender reports that a modified public tool read and decrypted credentials stored by the backup software. The attacker also copied a browser credential database and used Windows credential-theft tools.

This is evidence of credential abuse inside an already accessed environment. The report identifies no Veeam vulnerability or affected-to-fixed version range that would support a specific patch instruction.

The attacker created domain-administrator accounts named library and svc_bkpmon. On September 18, it configured two-way trusts between three previously separate Active Directory domains, the directories that manage accounts and computers. A trust lets accounts authenticate across domains; permissions still require separate grants. The attacker also pushed the backdoor account’s access to roughly 40 machines.

On September 19, it tried to remove endpoint security software, delete Windows system-state backups and deploy encryption. Bitdefender says anti-tampering controls blocked the observed security-agent removal attempts. About 37 machines were nevertheless encrypted. Blocking one attempted action did not contain the whole intrusion.

Keep backup access and domain administration separate

The report’s defensive recommendations follow the access paths in the logs:

  • Separate privileged accounts across domains. Avoid keeping credentials for several domains together. Put new domain trusts and domain-administrator accounts behind change control, and alert on unexpected additions.
  • Isolate backup infrastructure. Scope backup accounts to the domain and resources they need. Keep immutable recovery copies that ordinary domain accounts cannot delete.
  • Limit remote administration. Restrict Windows remote execution, remote registry access and administrative file shares to known management hosts. Review exceptions that let an ordinary workstation administer servers.
  • Close visibility gaps. Account for unmanaged hosts alongside protected endpoints. The unmonitored staging machine was both an attack source and a limit on the investigation.
  • Investigate credential access to backup systems. Compare unusual reads of stored credentials and Windows security data against known backup and administration jobs.

These recommendations address ordinary ransomware prerequisites. They do not depend on proving that a model wrote the attacker’s commands.

Separate observed behavior from the AI claim

Bitdefender points to repeated commands, numbered script revisions and retries spaced 40 to 101 seconds apart. Both tiny edits and larger rewrites arrived at similar intervals. That is consistent with an agentic loop, software that feeds command results back to a model and runs its next suggestion.

The investigators could not inspect the attacker’s machine. They could not identify a model or exclude a person copying results between a shell and a chat window. Their logs contain detected events only, so some attempts may be missing. The report explicitly calls this one case, rather than evidence of a wider trend.

Bill Toulas’s reporting summarizes the same investigation; it is not independent forensic confirmation. Treat the encryption and credential abuse as source-reported incident findings, and the degree of AI control as unresolved.

Caveat: Bitdefender sells endpoint protection, managed detection and threat intelligence; its published case cannot establish how autonomous the attacker was or how often similar attacks occur.