[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Threat Hunting · markdown

official (Bitdefender research, Nov 3). The AI involvement is Bitdefender's interpretation of command behavior; it says it could not see the attacker's machine or model

Hunt: a 3AM ransomware affiliate hit a UK school estate in what Bitdefender calls a part-human, part-agentic-loop intrusion: Veeam credential theft, rogue admins 'library' and 'svc_bkpmon', new two-way domain trusts

  • Backup first: a modified Veeam-Get-Creds.ps1 read and DPAPI-decrypted stored credentials on three Veeam servers via Impacket-style remote execution; a scheduled task named VEEAMPROBE was created on a backup server.
  • Credentials: reg save of SECURITY and SYSTEM hives; Mimikatz renamed sysmon64.exe, svcstat.exe and random .tmp.exe; a comsvcs.dll MiniDump of LSASS; Chrome Login Data copied; portable Python 3.11 running crackw.py from C:\Windows\Temp.
  • Persistence: domain admins library and svc_bkpmon, created by two different routes.
  • Unusual exfil: files staged as base64 registry values (HKLM\SOFTWARE\FIN001…FIN040, FINCOUNT, HKLM\SOFTWARE\VC31), zipped and POSTed to rcv.aspx on the unmanaged host, with alive HH:mm heartbeats.
  • Spreading: dnscmd /zoneadd … /dsforwarder then netdom trust … /add /twoway joined three separate AD domains; grant_lib.ps1 pushed by scheduled task to about 40 machines; av_remove.ps1/av_remove2.ps1 tried to strip Sophos and Bitdefender (blocked by anti-tamper).
  • Impact: host.exe -k <id> -s 15 with wbadmin delete systemstatebackup -keepVersions:0 -quiet.
  • Model-style artifacts worth alerting on: files and tasks named da_proof.txt, psexec-ok, TASK-OK, test123, step1.txt…step3.txt; output like ===TRUSTS=== or == CONNECTIVITY ==; 2>/dev/null inside cmd.exe lines; Impacket wmiexec output files __<timestamp> in ADMIN$.

Bitdefender publishes the full IoC set on its GitHub IoC repository. Its main point for defenders: an agentic loop makes an intrusion noisy and repetitive, not stealthy, so prevention controls that block an action the first time beat detections that rely on speed.

Sources: Bitdefender: 3AM Hits European Schools: A Hybrid AI Ransomware Attack (Nov 3) · BleepingComputer: UK school estate hit by 3AM ransomware in suspected AI-assisted attack