Test of ARTEX agents shows they make noise and claim success in error
In a test, an ARTEX agent said it had access to a Windows host when it had only a network connection to a service.
Reveal Labs, the research team of the security company ExtraHop, tested ScopeWeaver. This is a fork (a copy) of ARTEX, an attack tool that uses AI agents. In one test the agent declared success, but it had no remote shell, only a working connection. The tool made network traffic that the RevealX product of ExtraHop detected, also when the tool did not work.
How it works
A planner divides a goal into small intents. Workers operate tools at the same time and record results in a database. Each new plan uses these results. The agent used the scan tool Nmap, then PsExec and Evil-WinRM.
The numbers
In a second test, ScopeWeaver got a shell through PsExec of Impacket and turned off the Windows lock screen. It did not make a permanent access. Other groups report a wide use of ARTEX:
- Censys counted 1,465 exposed ARTEX consoles.
- Microsoft says ARTEX-aided login tries hit more than 2,250 Microsoft 365 tenants.
- Google saw bursts of two to 25 seconds.
What is not known
The setup tested how the tool operates. It did not show an attack with no help. The researchers say that user-agent strings and keep-alive signals do not prove the use of ARTEX alone. To read encrypted traffic, a team needs TLS decryption and session keys.
Insight: Agent attacks are easier to find by sequence than by tool
The FBI warned on 13 November that ARTEX works at machine speed. Google saw agents correct errors and reach goals in seconds, and Check Point found 1,608 live ARTEX panels. This test shows that the same speed has a weak point. An agent that tries again and changes tools leaves a trail, so defenders get more from a connected sequence of steps than from a search for AI. Expect vendors to sell sequence detection.
Sources
Posted