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

← Threat Hunting · markdown

vendor research (Malbear Labs IOCs and behaviors, Nov 8, 2026)

Hunt: Operation Snowball, a Sliver domain, PowerShell watchdog files in C:\Windows\Temp, Unicode-escaped JSP web shells and BladeLogic RSCD pivots

  • Network: Sliver C2 api.tablacysexplorer[.]com (Cloudflare-fronted, 443); 146.70.53[.]182:2299 (SSH to Sliver server); 113.23.212[.]15:8000 (tool staging); 212.46.38[.]73:56777 (operator machine); egress via VPNGate relays.
  • Windows: several .ps1 files in C:\Windows\Temp where a watchdog re-hashes a command script every 5 seconds and updates a lock file, plus results/done-flag files; two hidden PowerShell processes created via WMI Win32_Process.Create with -enc payloads connecting out on TCP 21; IFM/shadow-copy NTDS dumps followed by quick deletion.
  • Web tier: JSP files whose scriptlets are written as \uXXXX escapes (normalize before signature matching); Behinder JSP, WAR deployments via JBoss manager; ASPX shells next to services named like Update Orchestrator components; PHP shells taking ?c= with encoded PowerShell.
  • Lateral: connections to BladeLogic RSCD 4750 from application servers; Zabbix native-protocol (ZBXD header) queries from non-Zabbix hosts; SAP DIAG probes on 3220; Liferay portal_ext.log reads.
  • AD recon for AI assets: LDAP searches for hostnames containing OLLAMA, LLM, JUPYTER or GPU.

Sources: Malbear Labs: Operation Snowball: Inside an Autonomous AI Intrusion Campaign (Nov 8, 2026)