Censys Maps ARTEX Infrastructure and Related AI Attack Frameworks

· Original article ↗

Summary

Censys ARC reports identifying 1,465 ARTEX-associated hosts and web properties. Scan history and host pivots revealed exposed agent files, rapidly changing open directories, co-located threats and self-hosted LLM gateways.

Key points

  • CrowdStrike linked ARTEX to intrusions at South Korean financial institutions; Microsoft tied it to Storm-0940’s ConnectWise supply-chain compromise.
  • Censys scan history showed an ARTEX-associated directory serving files from August 18, 2026, including a .claude folder, prompts, session histories and configuration files.
  • Censys says its ARTEX label covered 1,465 hosts and web properties; 155 also had another detected threat, and 43 ran self-hosted LLM gateways.
  • ARC counted 32 open directories on ARTEX hosts. From November 3 to 5, ten closed and fifteen new ones appeared.
  • Censys found 70 hosts shared by ARTEX and CyberStrikeAI, and tracked related frameworks including ScopeWeaver, PentAGI, HexStrike AI, RedAmon and Cairn.
  • The article recommends using scan history, exposed agent artifacts and console fingerprints for investigation and detection; it notes that some exposed consoles may belong to authorized testers.

Article Details

Attack Vectors
  • ARTEX is described as an autonomous, multi-agent AI penetration-testing and offensive framework that was used in reported intrusions.
  • An ARTEX operator exposed a working directory through an open directory on a public host, revealing prompts, session histories, and configuration files.
  • The article says operators may expose working directories using short-lived Python HTTP servers on high ports, making agent instructions and configuration files publicly accessible.
  • Censys observed attack-framework consoles and other detected threats co-located on some hosts; the article cautions that some consoles may belong to authorized testers.
Defensive Notes
  • Continuously monitor exposed Internet-facing consoles and services, and close those that should not be public.
  • Investigate IPs and domains even when reputation services show no hits; a clean reputation result does not establish that infrastructure is benign.
  • Use scan history to investigate what a host served during an intrusion, including directories that have since closed.
  • Monitor for agent artifacts such as CLAUDE.md, AGENTS.md, GEMINI.md, .claude directories, and .mcp.json files exposed through open directories.
  • Use console fingerprints and deployment signals, including port 8787, unfamiliar LLM gateways, mitmproxy banners, and short-lived Python HTTP servers on high ports, to identify suspicious hosts and earlier activity.
  • Review outbound connections for unfamiliar LLM gateways and API resellers, and check organizational IP ranges for exposed attack-framework consoles.

Indicators of compromise

TypeIndicatorContext
IPV438[.]244[.]50[.]120CrowdStrike identified this host as the location of an ARTEX operator's open directory, which exposed prompts, session histories, and configuration files.
IPV443[.]255[.]159[.]10Censys scored this host as malicious, despite vendors not flagging its IP or associated DNS resolution.

Threat Actors

Vendors

Products

Tools

Countries

Related Articles