AI Security

ARTEX AI Pentesting Framework: A Defender’s Field Guide

Photo: Lenharth Systems (CC0 1.0)
In this article5 sections

What ARTEX is

ARTEX is an open-source, LLM-driven autonomous penetration-testing framework: an operator hands it a target and an objective in natural language, and a team of LLM agents plans and runs the attack with real tools rather than simulated actions. CERT-EU documented that architecture in an advisory published on 5 November 2026, and the short version for defenders is that a sentence now compiles into a running intrusion plan.

The design detail that matters most is that ARTEX ships no model. Datadog Security Labs reports that operators supply API keys for OpenAI-compatible, OpenAI Responses or Anthropic Messages endpoints, which lets it drive Claude, GPT, DeepSeek, Qwen, GLM, Kimi or a self-hosted model of the operator’s choosing — and move to a different provider when a key is cut off. There is no single vendor endpoint you can block to make the framework stop thinking.

Its release history is messy. The developer published it on GitHub in July 2026, according to Datadog Security Labs, while the FBI FLASH alert places the publication in August. The repository was removed on 8 October 2026 after reports of misuse, by which point it had more than 12,000 forks. Removal did not remove the capability: forks, archived copies and a commercial version keep it circulating.

Under the hood it is deliberately unremarkable. Microsoft and Datadog both describe default tooling built from Nmap, curl, wget, Python HTTP clients and headless Chromium driven through Playwright. Nothing on that list is exotic, and most organizations already have every one of those tools installed somewhere. The novelty is the planner on top, not the payloads underneath.

The framework is also not rare in the wild. Censys counted more than 1,600 publicly exposed ARTEX console instances, and Check Point Research reported more than 1,800. Both figures are lower bounds, because private deployments are invisible to internet-wide scanning.

What attackers are doing with it

The FBI FLASH alert of 13 November 2026 states that ARTEX can “automate[s] multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration,” and notes that some intrusions reached privileged access from initial targeting in under three hours. Read that in terms of your own detection coverage: most alerting assumes an attacker who is slow, noisy and sequential.

Microsoft, in a report dated 6 November 2026, attributes observed use to Storm-3358, a Cambodia-based, financially motivated cluster hitting financial organizations in Southeast Asia and a smaller number of large global financial firms since October 2026. The most useful part of Microsoft’s account is what sits alongside the framework: anonymous file upload and server-side request forgery exploitation, web shells, and the WISP, vshell and AdaptixC2 remote access trojans downloaded with netcat and curl.

That pairing is the pattern to internalize. The agents handle reconnaissance, triage and planning — deciding which of your exposed services is worth an hour of effort. The exploitation and post-exploitation payloads stay commodity. An ARTEX intrusion therefore looks, at the endpoint and network layer, a great deal like an ordinary web-application compromise, with the enumeration widened and the planning compressed.

The victim reporting is concrete. BleepingComputer reported that Seoul police linked a group to more than 90 breached organizations, with ransom demands of 360 million won (about USD 270,000) in 38 cases. The Wall Street Journal reported that 68,000 people had data exposed. For small and mid-sized organizations outside finance, the takeaway is not “we are not the target” — it is that the cost of finding out you were one went down for the attacker and up for you.

How to tell it is running in your environment

You will not get a signature. What you get is a legitimate toolkit, driven at a pace and in an order that humans rarely manage. Build detections around those two properties — the tools and the tempo — and treat every hit as a lead that needs corroboration rather than a verdict.

Start in the web tier, where the framework does most of its reconnaissance, then follow it onto the host and out to the network edge.

  • User agents that name the framework. Microsoft and Datadog both note observed user agents that literally contain ARTeX, including “Mozilla/5.0 ARTeX”, “artex-pentest/1.0” and “ARTEX-worker/1.0”. Add a WAF and log-search rule for the string, alert on it, and keep the rule live for months — it is trivially spoofable, so treat a hit as a strong indicator rather than proof, and never as the only rule you rely on.
  • Headless browser fingerprints. Default tooling includes headless Chromium via Playwright. Look for browser-shaped traffic with no matching human behavior: instant form fills, no mouse or scroll telemetry, consistent viewport and language settings across sessions, and one client touching many unrelated application paths.
  • Enumeration bursts from a single source. Nmap, curl and Python HTTP clients hitting directory lists, API routes and parameters in quick succession from the same address or ASN, sitting on a 404-heavy baseline.
  • Shell children you did not expect. A web server or application service account spawning curl, wget or netcat is the highest-value alert in this article, because Microsoft lists exactly those tools as the download mechanism for the WISP, vshell and AdaptixC2 trojans. Baseline normal behavior first, then alert on the delta instead of switching it on cold and drowning.
  • Upload-then-fetch chains. Correlate a POST to an anonymous upload endpoint with a following request from the server to internal address space, loopback or cloud instance metadata addresses. That is the upload-plus-SSRF pattern Microsoft describes, and it is cheap to express as a correlation rule if proxy and application logs already share a platform.
  • New files in web-accessible directories. Extend file integrity monitoring to document roots and upload directories, not just system paths. A web shell written into a web root is the quiet step between initial access and the RAT download.
  • Exposed consoles and management planes. Censys and Check Point Research both found large numbers of publicly reachable ARTEX consoles, and both counts are floors. Inventory your own internet-facing management interfaces — agent consoles, admin panels, orchestration UIs — and confirm which ones genuinely need to face the internet.

Then add the signal that ties them together, because single-event alerting will not catch a kill chain that completes in under three hours. Track phase progression per source: reconnaissance, then authentication attempts, then a successful login, then an upload, then outbound transfer, all inside one shift. An alert that fires on the fourth phase from the same source in a short window beats four independent alerts nobody correlates at 3 a.m.

What to do this week

CERT-EU’s defensive advice is short enough to work through in a week and specific enough to assign to an owner. Sequence it by what an autonomous agent exploits fastest.

  1. Cut internet exposure to the bone. Every service you retire from the public internet removes a candidate from an agent’s reconnaissance list, and reconnaissance is the phase ARTEX compresses hardest.
  2. Isolate management planes. Administrative interfaces belong behind a jump host, VPN or identity-aware proxy. That includes the consoles and orchestration UIs you built yourself and then forgot about.
  3. Emergency-patch internet-facing systems. Do not wait for the next maintenance window. Prioritize by reachability and exploitability rather than CVSS alone; the goal is to shrink the set of chainable findings.
  4. Monitor privileged accounts and enforce MFA. Enforce it everywhere, and assume the session token, not just the password, is the prize. Pair the rollout with defenses tuned for adversary-in-the-middle phishing, such as the token-theft patterns covered in our guide to defending against AiTM phishing with N0va Phishkit.
  5. Hunt for known infrastructure. Search logs for the user-agent strings above, for exposed-console fingerprints and for the tooling combinations Microsoft published. Promote each hunt into a permanent rule where it holds up.
  6. Keep low-severity findings in the remediation cycle. This is CERT-EU’s most underrated recommendation. An agent does not get bored chaining a medium-severity SSRF into a file upload into a web shell. The findings you deprioritized are the ones it will reach for.
  7. Write the three-hour assumption into your response plan. If privileged access can arrive before your on-call engineer finishes triage, pre-arranged capacity matters more than a contact list. Understand what the clock actually covers by reading what an incident response retainer pays for.

What this changes about AI risk

The important shift is not that AI can hack. It is that the expensive human bottleneck — deciding what to try, reading the output, adapting when an attempt fails — has been absorbed by the model layer, while execution stayed with ordinary tools. That is the same trajectory traced in our broader look at AI and machine learning in cyber security: capability arrives as automation of judgment, not as a new class of exploit.

Two consequences follow. First, breadth now beats depth for the attacker. An agent can probe fifty services in the time a human spends on one, so your weakest internet-facing asset defines your exposure, not your best-defended one. Second, the low-severity backlog became offensive infrastructure. Every “we will get to it” finding is a step an agent may take on your behalf.

There is an unglamorous flip side. Your own organization now runs agents holding credentials and browser sessions. The same headless-browser automation that makes ARTEX effective is what makes an AI assistant a valuable foothold once its extensions or session tokens are compromised, which is why defending AI assistants from malicious browser extensions has moved from nice-to-have to priority.

What does not change is the fundamentals. Patching, segmentation, MFA, least privilege and log correlation still decide the outcome; ARTEX raises the price of skipping them. The teams that come through this well will have treated the framework as a tempo problem rather than a mystery — and spent the week closing the findings they had already decided were not urgent.

Nathan Cole

Vulnerability management research, Dominion Cyber

Nathan Cole writes about vulnerability management and emerging threat research — why CVSS score alone is a poor prioritisation input, how patch operations actually get run, the mechanics of adversary-in-the-middle phishing, and the security model of AI assistants and browser extensions.

Get the weekly security brief

One email a week: what is worth patching, what is worth watching, and what is worth reading. No spam, unsubscribe any time.