FBI on ARTEX: ATT&CK techniques, new indicators and defenses
FBI FLASH on ARTEX: observed ATT&CK techniques, new IPs from the ConnectWise and Storm-3358 campaigns, LLM API traffic signals and recommended defenses.

On 13 November 2026 the FBI published FLASH 2026-1113-001 (TLP:CLEAR) on ARTEX, the LLM-driven autonomous penetration testing framework that ARX News has already covered in connection with attacks on South Korean banks and Indian fintechs. The advisory brings together analyses by Microsoft, Google Threat Intelligence Group and Mandiant, CrowdStrike, Zscaler, Censys, Check Point, Akamai and CERT-EU, and adds three things defenders can use: a mapping of the observed activity to MITRE ATT&CK, new network indicators and a list of recommendations. Everything that follows comes from the FBI advisory, the only source for this article.
The picture according to the FBI
- Since at least August 2026 multiple malicious actors have used ARTEX against organizations worldwide: financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications and legal. In some intrusions it took less than three hours to go from initial targeting to privileged access.
- ARTEX is written in Go and is multi-agent: it turns goals expressed in natural language into commands executed by specialized agents. It has no model of its own and connects to external providers. On attacker-controlled instances, Chinese-origin models dominate, above all the DeepSeek Flash family, because of cost and the lack of observable guardrails when requests are framed as penetration testing.
- Released on GitHub in July 2026 by the developer “Autumn-27”, the repository was made unavailable on 8 October; the code keeps circulating through forks, variants, archived copies and a commercial version, ARTEX PRO.
- In early November researchers counted more than 1,600 ARTEX consoles exposed to the internet; according to the FBI, total deployments are likely much higher.
- Campaigns cited include the South Korean financial institutions, with personal data of about 67,000 people, and the supply chain compromise of ConnectWise Cloud Backup, which according to the FBI enabled access to more than 2,250 downstream tenants. Agents also modified iOS app distribution infrastructure using exposed cloud credentials.
Observed techniques
The advisory maps the activity by phase. In short:
- Reconnaissance: automatic mapping of the effective permissions of Azure credentials (SAS tokens, service principals, Kudu credentials, storage account keys) with sub-second checks, test blob writes and deletes and ARM deployment validation dry-runs; IAM permission tests using DryRun flags; web directory fuzzing with more than 500 requests (T1595, T1580, T1087.004, T1526).
- Initial access: unsigned SAML assertions, XXE, SSRF, SQL injection, SSTI, command injection, IDOR, unauthenticated web APIs, administrator passwords hardcoded in single-page application JavaScript bundles, stolen VPN credentials, cross-tenant cloud trust relationships and vendor B2B federation (T1190, T1078, T1199, T1195, T1606.002).
- Persistence: JSP and PHP web shells, including via MySQL log poisoning; commands through Kudu administrative APIs; open-source RATs WISP, vshell and AdaptixC2; credentials added to existing OAuth applications (T1505.003, T1098.001).
- Escalation and internal reconnaissance: Kerberoasting, DCSync for the krbtgt key, abuse of resource-based constrained delegation, cloud metadata read via XXE, administrator tokens forged with unrotated keys; enumeration of more than 110,000 Active Directory objects, reading LAPS passwords and inventorying security sensors to find Windows machines without CrowdStrike Falcon (T1518.001).
- Lateral movement: attempts with Pass-the-Ticket, Pass-the-Hash, WinRM and Kerberos relay against domain controllers, blocked by endpoint protection.
- Collection and command and control: mass blob and database reads, 100 GB volumes created to stage exfiltration, command output sent to rotating webhook.site endpoints; traffic through residential proxies, Tor, SOCKS, Chisel and SoftEther tunnels; the BLOODALCHEMY malware communicating with C2 domains (T1090.003, T1572).
The FBI notes that ARTEX generates a lot of noisy traffic, but its speed leaves little time for a coordinated response.
New indicators
In addition to those already known from the Korean campaign (38.244.50[.]120, the nine proxies and the API reseller xcai[.]pro), the advisory lists:
- ConnectWise campaign (source: Microsoft): 144.172.64[.]60, 144.172.64[.]99, 144.172.64[.]128; for Graph access through compromised apps, transient VPN IPs relevant on 15–16 October 2026: 187.15.88[.]186, 187.15.90[.]217, 187.15.89[.]186;
- Storm-3358 campaign (source: Microsoft, active as of 6 November 2026): WISP C2 nodes 199.19.95[.]66, 199.19.95[.]67, 199.19.95[.]68; ARTEX servers 91.92.240[.]238, 85.11.168[.]45, 178.16.54[.]129, 31.77.229[.]35; C2 server 85.11.167[.]52.
Behavioral signals (source: Zscaler): unexpected calls to api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai alongside other suspicious activity; bursts of requests with sequential enumeration against login and search pages; .aspx or .jsp files in unexpected web paths. The FBI recommends vetting the indicators before blocking them: cloud IPs may have been reassigned.
Impact for European and Italian companies
Energy, healthcare, manufacturing, telecommunications, finance and public administration are NIS2 sectors, and CERT-EU is among the advisory’s sources. The entry points described are not exotic vulnerabilities but common mistakes: partner and supplier portals, mobile backends, applications built by individual business units, secrets in code, exposed cloud management interfaces. With timelines of a few hours, the response has to start from alerts that are already in place.
What to do now
From the FBI recommendations:
- Inventory all internet-facing applications, peripheral ones included. Restrict cloud administration, Kudu/SCM endpoints, hypervisors and backup platforms to trusted networks, private connectivity or dedicated administrative paths; where feasible, move enterprise applications behind Zero Trust access.
- Disable basic publishing authentication where it is not needed and require an architecture review before exposing a service on a public IP.
- Identity: least privilege for multi-tenant apps and consent grants; in Entra ID review credentials added to applications, new service principals and owners, privileged role assignments and Conditional Access changes; review successful and failed service principal sign-ins separately.
- Applications: rate limits and thresholds on business APIs, alerts on sequential lookups in portals, server-side authorization, no stack traces or SQL errors for unauthenticated users.
- Secrets: continuous scanning of repositories, container images, pipelines and infrastructure-as-code, with prompt rotation of exposed credentials.
- ConnectWise customers: review every Microsoft 365 tenant where the apps are or were authorized, searching by application ID; review activity from at least 1 October 2026; keep the connections disabled until ConnectWise issues formal restoration instructions and preserve evidence before deleting enterprise applications.
How ARXDOME helps
ARX Expose shows what an attacker sees from outside, so a forgotten partner portal or management interface surfaces before an automated agent finds it. ARX Horizon enriches events with threat lists, and ARX Sentry brings server and firewall logs into the SOC and NOC views, where you can look for the published addresses and for connections to language model APIs. Checks in Entra ID and Microsoft 365 tenants still have to be done in their admin centers.
Last updated:


