ARTEX Detection: Hunting the AI Pentest Framework Turned Attack Platform

ARTEX Detection: Hunting the AI Pentest Framework Turned Attack Platform

SOC Prime Team
SOC Prime Team linkedin icon Follow

Offensive AI is no longer a slide in a threat-landscape deck. It is sitting on attacker servers, wired to multiple LLM APIs, and probing production APIs at machine speed. ARTEX, an open-source agentic AI penetration testing framework developed in China, is the clearest example so far: a red-team tool that threat actors have adopted to automate reconnaissance, vulnerability probing, and exploitation end to end.

For SOC analysts and detection engineers, ARTEX detection is a new kind of problem. There is no single malware family to fingerprint. Instead, you face an LLM-driven operator that adapts payloads, rotates techniques, and generates traffic volumes no human pentester could. This article walks through the ARTEX analysis available so far, the behaviors worth alerting on, and how to detect ARTEX across web, endpoint, and Linux telemetry.

Get Started For Free

ARTEX Analysis

ARTEX went public on GitHub as Autumn-27/ARTEX on July 26, 2026, under AGPL-3.0. It is written in Go, runs a Next.js console on PostgreSQL, and sits on top of ScopeSentry, the same author’s asset-discovery platform (Censys).

The architecture explains the telemetry you will see. A Planner agent turns an operator goal, such as “gain admin panel access,” into attack intents. Worker agents claim those intents, run tools like nmap, and write results back into an asset graph and an exploration graph. The loop repeats until the goal is met, using Anthropic, OpenAI, or any OpenAI-compatible model endpoint.

That loop is fast. Mandiant and GTIG recorded 27 events within 23 seconds, plus automated error correction used to land webshells, as covered in our report on how ARTEX leads a new wave of agentic offensive frameworks alongside PentAGI and CAIRN.

The author disowned malicious use on October 8 and took the repository down, but forks and rebrands such as ScopeWeaver still circulate. As of early November, Censys counted 1,465 exposed ARTEX consoles, most on default port 8787.

ReportedCampaignWhat ARTEX affects
Nov 9, 2026Storm-3358 targets financial organizationsSSRF and anonymous file uploads, then WISP, vshell, and AdaptixC2 RATs plus cloud identity abuse
Oct 27, 2026Storm-0940 hits IT and software vendorsOkta, Entra ID, and Cognito auth probing; account recovery abuse; IP and user-agent rotation
Oct 21, 2026ARTEX found on servers attacking Japanese companiesAPI probing and WAF evasion in Japan and South Korea, driven by DeepSeek, Zhipu AI, and xAI models
Oct 21, 2026Storm-0940 ConnectWise supply-chain compromise (Microsoft)Sign-ins against 2,250+ tenants, peaking at 2,900 attempts per minute
Oct 7, 2026South Korean banks (CrowdStrike)A broker-facing loan service and an employee mobile system breached, data exfiltrated

Detect ARTEX

Detect ARTEX by behavior, not by name. Binaries get renamed and consoles get rebranded, but the explore-decide-execute loop leaves volumetric and sequential fingerprints across five log sources.

  • Web server: 40x response spikes from a single IP, path traversal, file inclusion, and OS command injection patterns, and scripted files dropped into web directories. Wiz saw a fetch of /swagger/?format=openapi followed immediately by a random nonexistent URI, a clean mapping signature.
  • Identity: failed-login bursts and access-denied spikes on Okta, Entra ID, and Cognito, password spraying, and abuse of account recovery workflows.
  • Endpoint and Linux: web service accounts spawning shells, SSH known_hosts discovery, nmap, curl, or wget from unexpected parents, and netcat paired with RAT binaries.
  • Cloud: an Azure write-then-delete of a throwaway blob, and AWS EC2 calls with –dry-run that test IAM permissions without changing state.
  • Network: HTTP C2 to direct IPs on uncommon ports, and egress to LLM resellers or self-hosted proxies such as CLIProxyAPI and New API.

ARTEX IOCs worth loading now include the ARTEX host 38.244.50[.]120 and LLM reseller xcai[.]pro (CrowdStrike), plus Storm-0940 infrastructure 144.172.64[.]60, .99, and .128 (Microsoft). Treat them as perishable: Censys watched 10 attacker open directories close and 15 new ones appear in two days.

For ARTEX mitigation, rate-limit by account, session, and API key rather than IP alone, enforce API authorization server-side, and issue short-lived credentials. Decoys and tarpits work unusually well here, because an autonomous agent will happily burn its token budget on fake assets. Every Active Threats report linked above ships ready-to-deploy Sigma and AI rules plus simulation scripts, so you can validate coverage before ARTEX tests it for you.

PHP2shell ARTEX

PHP2Shell is the bug class an agentic framework loves most: pre-auth, widespread, and with public PoCs. CVE-2026-109618 and CVE-2026-109619 are RCE chains in PHP core’s session.upload_progress, rated CVSS v4 9.2 and affecting PHP 5.4.0 through 8.5.11. PHP shipped fixes on October 21, 2026, a day early after the CVE IDs leaked (Arctic Wolf).

The ARTEX PoC angle is concrete. An issue on an ARTEX fork proposed a php2shell-check skill that vendors three public verification scripts from the php2shell-poc research repo. Once a check like that exists as a skill, the Planner can queue it as just another intent against every PHP asset in the graph.

The exposure is huge. Qualys counted 167 million PHP installations, 71% of them on end-of-life versions that will never get a vendor patch (Qualys). Amazon caught an exploitation attempt in its MadPot honeypots on October 17, four days before disclosure.

To hunt PHP2Shell in ARTEX-style traffic, look for:

  • Multipart POSTs carrying a PHP_SESSION_UPLOAD_PROGRESS field with attacker-chosen session IDs
  • Follow-up file inclusion requests pointing at sess_ files in PHP session directories
  • Inert PHP canaries and unique output markers, which scanners use to confirm evaluation
  • PHP worker processes spawning shells, curl, or wget

Patch first, upgrade end-of-life branches, and disable upload progress tracking where applications do not need it.

FAQ

What is ARTEX and how does it work?

An open-source, LLM-driven pentest framework. Planner and Worker agents loop through recon, probing, and exploitation toward an operator-defined goal.

When was ARTEX first discovered?

It was released July 26, 2026. CrowdStrike first reported malicious use on October 7, 2026.

What is the impact of ARTEX on systems?

Data theft, RAT deployment, cloud identity takeover, and supply-chain compromise across finance, IT, and software vendors.

Can ARTEX still affect me in 2026?

Yes. Forks and rebrands remain active, and Censys still tracks 1,400+ exposed consoles.

How can I protect myself from ARTEX?

Patch fast (PHP2Shell first), rate-limit by identity, harden IdP flows, and deploy decoys.

How can I detect ARTEX?

Correlate volumetric web, identity, and cloud anomalies. Full details for ARTEX rules live in the SOC Prime reports linked above.

Join SOC Prime's Detection as Code platform to improve visibility into threats most relevant to your business. To help you get started and drive immediate value, book a meeting now with SOC Prime experts.