Skip to main content

ExtraHop Tests ARTEX Fork: AI Attack Agents Claim False Success

ExtraHop's Reveal Labs tested ScopeWeaver, an English-language fork of the ARTEX agentic attack framework, in a lab network and published the results on 15 November 2026. The agents mapped the network and tried many remote-access tools, but in one run claimed success without getting a shell. ExtraHop lists user agents and traffic patterns defenders can watch for.

On this page

What ExtraHop tested

ARTEX is an open-source, multi-agent penetration testing framework that a developer published on GitHub in July 2026 and removed on 8 October, after attackers began using it; forks and copies are still available. Microsoft says a China-aligned group it tracks as Storm-0940 used ARTEX in a campaign that tried to sign in to more than 2,250 Microsoft 365 tenants and fed into the ConnectWise supply chain compromise. Microsoft later reported a Cambodia-based group using it against financial firms.

On 15 November, ExtraHop's Reveal Labs published what happened when it ran ScopeWeaver, an English-localized fork of ARTEX, against a lab network. Neither Microsoft report names ScopeWeaver, and no source ties this specific fork to the real-world campaigns. The value of the research is a close look at how these agent tools behave on a network.

How the agents work

ExtraHop describes a planner agent that breaks a goal into smaller intents, then hands them to worker agents running in parallel. Workers run tools, write what they find to a PostgreSQL database, and that store feeds the next round of planning. Microsoft's analysis of ARTEX lists Nmap, curl, wget, Python HTTP clients and headless Chromium among its default tools, and says the framework can use any AI model.

ExtraHop adds that ScopeWeaver can use locally hosted language models or cloud APIs, falling back from one model to another. That matters for defenders: an attacker running an open-weight model on their own hardware leaves no trail at an AI provider, so provider-side abuse detection cannot be relied on to stop them.

The agent said it won when it had not

The lab goal was to reach a Windows host and then get a shell on a domain controller. In one run, SSH and RDP attempts failed, Nmap found the WinRM ports, and the agent cycled through PsExec, xfreerdp, rdesktop and Evil-WinRM without getting in. A connectivity check worked, but no shell opened. The agent still reported success.

In a separate run, Impacket's PsExec did give it a shell, and the agent changed a lock screen setting but set up no persistence. The pattern is the same weakness seen in other AI agents: they can be fast and persistent while still misreading their own results, which cuts both ways for the people running them.

Signs defenders can look for

ExtraHop lists user agent strings seen in testing, including "scopeweaver" and "artex-enrich/1.0", along with function names such as list_findings, record_fact and shell_send in model API traffic, and idle workers sending keep-alive messages about every five seconds. It cautions that none of these alone proves malicious ARTEX use. It also cites counts of more than 1,400 exposed ARTEX consoles online.

Microsoft's advice goes further than blocking signatures, which it says will not stop an agent on their own. It suggests watching for bursty, retry-heavy traffic and default tool user agents, and using honeypots, decoys and fake wins to steer agents away from real systems. ExtraHop's false-success finding suggests why decoys can work: an agent that cannot reliably tell success from failure is easier to mislead.

For organizations, the practical steps are familiar: patch exposed services, limit WinRM and remote admin tools to the hosts that need them, review multi-tenant app permissions, and alert on unusual scanning from inside the network.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.