Skip to content
ARTEX.sh

How to detect ARTEX:speed, volume, LLM traffic and published markers.

AI agent attacks are loud but fast. This page collects the signs that vendors and agencies have published for ARTEX activity, so a security team knows what to watch for. It is detection guidance only.

Checked

Watch behavior, not only indicators

IP addresses and file names change fast. The FBI says its listed indicators should be checked before blocking and may not mean a compromise on their own. Behavior is harder to change: an agent works at machine speed, tries many small variations, and runs several paths at once.

Elastic Security Labs advises hunting request tempo per identity and joining it across sessions, so that many short sessions from one account still add up.

Speed and volume

The FBI describes these patterns in ARTEX-driven attacks:

  • High-frequency automated requests that walk through values in order against login and search endpoints.
  • Short, dense bursts of activity.
  • Many failed attempts, each changed a little from the last.
  • Signs of several agents stepping on each other, such as duplicate requests.
  • Unusually many records read per session on partner or employee portals, and rate limit hits on APIs.
  • Mass emails or notifications sent from a compromised application.

LLM API traffic where it does not belong

ARTEX needs an AI model to think. If an attacker runs it from inside your network or cloud, that traffic can show up in your egress logs. The FBI lists outbound calls to api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai as worth investigating when they appear next to other suspicious activity on systems that should not call them.

It also lists self-hosted LLM gateway or reseller software such as CLIProxyAPI and New API, unfamiliar LLM proxies in egress logs, and odd token use or API bills. Check Point and the FBI both say Chinese models, mostly the DeepSeek Flash family, were the most common in attacker-run instances.

Signs on hosts and in the cloud

These come from vendor reports and the FBI alert. Each is a lead to check, not proof on its own.

  • Command lines with artx or artex strings, cloud roles with an artx prefix, and the C:\ProgramData\artx5v\ directory (CrowdStrike, VAPOR PANDA report).
  • Web shells: .aspx or .jsp files in unexpected web app paths (FBI).
  • Outbound HTTPS transfers of structured data such as names, phone numbers or financial details to unknown destinations (FBI).
  • Cloud metadata queries, unexpected IAM role assumptions or PassRole calls, and test writes and deletes in cloud storage a fraction of a second apart (FBI).
  • New credentials on OAuth apps, new accounts with naming patterns, new SSH keys and unexpected new VMs (FBI).
  • The user agents artex-enrich/1.0 and norma/0.4 in web logs (StepSecurity).

Exposed ARTEX pentesting consoles: ports 8787 and 8788

The ARTEX web console and API listen on port 8787 by default. The recording proxy uses port 8788 and is bound to localhost in the standard Docker Compose setup, so 8787 is the port usually seen from outside. StepSecurity notes the supplied Compose file opens 8787 on every network interface.

Several teams counted exposed consoles. Censys found 1,465 hosts, about 815 distinct, on 2026-11-04. Check Point counted 1,608 live systems, and GitGuardian counted 1,428 ARTEX-family services on Shodan. GitGuardian publishes fingerprint descriptions so defenders can find their own instances, not as a target list.

Check your own address space first. An ARTEX console on your network that nobody approved is a finding by itself. A console somewhere else is not proof of bad intent: Censys says many belong to authorized testers.

Published alerts and lists to check

  • FBI FLASH-20261113-001: indicators, MITRE ATT&CK mapping and a list of what to send the FBI.
  • JPCERT/CC alert JPCERT-AT-2026-0030, case E: suspicious source IP addresses with observation dates. JPCERT/CC asks organizations to check their logs for them.
  • CrowdStrike's VAPOR PANDA report: the host and cloud markers above.
  • Censys: open directories on ARTEX hosts that exposed post-exploitation kits. Operators who leave these open expose their own tools.
  • DIVD's account of its own breach, summarized on the DIVD incident page: job labels made of an i and a number, and an admin API token with no expiry.

Questions

How can I tell if ARTEX is attacking my site?
Look for fast, automated bursts against login and search pages, many failed tries that change slightly each time, and too many records read per session. Then check the published markers on this page.
What port does ARTEX use?
Port 8787 for the web console and API, and port 8788 for the recording proxy. The proxy is bound to localhost in the standard Docker Compose setup.
Can I block ARTEX by IP address?
Only partly. Published IP lists help you check past logs, but the FBI says such indicators may be stale. Use them to hunt, and rely on behavior and rate limits to stop new attempts.
Does ARTEX leave a user agent?
StepSecurity lists artex-enrich/1.0 and norma/0.4. User agents are easy to change, so treat a match as a lead, and the absence of one as nothing.
Is LLM API traffic a sign of an attack?
Not by itself. Many companies use AI services. It matters when a server that should never call an LLM API starts doing so, together with other odd activity.

Sources

Every claim on this page comes from these sources or from this site's own sourced pages. Open them for the full reports.

Follow the ARTEX story.

New forks, releases and reports, dated and sourced.

Isometric plane with glowing light columns