← Voltar ao feed

MalwareMustDie releases open ARTEX detection pack (v1.1.3) for AI pentest framework abused by attackers

1 fonte(s) · 1 atualização(ões) · publicado 16/11/2026 08:50 · coletado há 1 dia

The researcher MalwareMustDie published a public, behavior-first detection pack for ARTEX, an open-source LLM-driven pentest framework that the article says criminals now run against real targets. The pack ships in seven formats: Sigma (five behavioral correlation rules), Sentinel KQL, Suricata 7, Snort 3, osquery, YARA and a STIX 2.1 bundle. IOCs expire after a year, while the behavioral rules do not. The article says the author updated the pack to v1.1.3 after traffic tests, added data from an FBI flash report, and confirmed previously held Sophos X-Ops indicators. This record rests on the article alone. No web research was done, so the report, pack and vendor statements were not checked.

Detalhes técnicos

  • ARTEX is an LLM harness, so changing the model, prompts or infrastructure changes the IOCs. The rules therefore target agent behavior: machine-speed recon chains, password sprays of about 1 attempt per second with rotating user agents, self-correcting API retries, and systematic record enumeration.
  • The playbook was tested on four campaigns. One was the October 2026 ConnectWise Cloud Backup case, where a compromised vendor app credential let service principals assign themselves directory roles. One became Global Administrator from a Microsoft-internal private IPv6 address. Attackers planted secrets that expire in 2028, and Microsoft said its own systems were not compromised.
  • The rule for the ConnectWise app ID now fires on role self-assignment or credential addition regardless of source IP.
  • A Korean financial case involved about 25,000 records pulled over roughly 30 hours (about 14 a minute). The pack counts distinct records read per identity over time.
  • A Japanese wave, reported by JPCERT/CC and GMO, used known bugs and misconfigurations and no zero-days.
  • An experimental Zeek approach is included. It cannot see Entra ID sign-ins.

Remediação

  • Deploy the behavioral rules and alert on identity actions, not source IP.
  • Treat generic indicators such as a curl user agent as enrichment only.
  • Test the pack and report false positives and true positives.

Linha do tempo

  1. securityaffairs.com ↗ · pub. 16/11/2026 08:50 · coletado 16/11 10:00 primeiro registro

    MalwareMustDie releases open ARTEX detection pack (v1.1.3) for AI pentest framework abused by attackers

    ler a análise desta fonte

    The researcher MalwareMustDie published a public, behavior-first detection pack for ARTEX, an open-source LLM-driven pentest framework that the article says criminals now run against real targets. The pack ships in seven formats: Sigma (five behavioral correlation rules), Sentinel KQL, Suricata 7, Snort 3, osquery, YARA and a STIX 2.1 bundle. IOCs expire after a year, while the behavioral rules do not. The article says the author updated the pack to v1.1.3 after traffic tests, added data from an FBI flash report, and confirmed previously held Sophos X-Ops indicators. This record rests on the article alone. No web research was done, so the report, pack and vendor statements were not checked.

    Technical Details

    • ARTEX is an LLM harness, so changing the model, prompts or infrastructure changes the IOCs. The rules therefore target agent behavior: machine-speed recon chains, password sprays of about 1 attempt per second with rotating user agents, self-correcting API retries, and systematic record enumeration.
    • The playbook was tested on four campaigns. One was the October 2026 ConnectWise Cloud Backup case, where a compromised vendor app credential let service principals assign themselves directory roles. One became Global Administrator from a Microsoft-internal private IPv6 address. Attackers planted secrets that expire in 2028, and Microsoft said its own systems were not compromised.
    • The rule for the ConnectWise app ID now fires on role self-assignment or credential addition regardless of source IP.
    • A Korean financial case involved about 25,000 records pulled over roughly 30 hours (about 14 a minute). The pack counts distinct records read per identity over time.
    • A Japanese wave, reported by JPCERT/CC and GMO, used known bugs and misconfigurations and no zero-days.
    • An experimental Zeek approach is included. It cannot see Entra ID sign-ins.

    Remediation

    • Deploy the behavioral rules and alert on identity actions, not source IP.
    • Treat generic indicators such as a curl user agent as enrichment only.
    • Test the pack and report false positives and true positives.