Microsoft says it has tracked what looks like the first large-scale campaign where the grunt work of hacking wasn't handled by a person at all, but by an autonomous AI agent. Behind it: Storm-0940, a China-linked group that has been stealing data from organizations worldwide since 2021.
The tool was ARTEX, a multi-agent framework written in Go with a Next.js front end. It was originally built for legitimate penetration testing — point it at a target, and it works through everything from reconnaissance to data exfiltration with barely any human input. Its developers took the code closed-source after ARTEX turned up in attacks on South Korean financial institutions. That didn't stop Storm-0940 from using it.
The way in was a ConnectWise cloud backup application, compromised on October 15 and disclosed publicly the next day. Using stolen application credentials, the attackers tried to log into more than 2,250 Microsoft 365 tenants — with just four stolen credentials, the agent fired off over 200 sign-in attempts in under four minutes.
- 151,000+ failed login attempts between October 10 and 14;
- a peak of 2,900 attempts per minute on October 11;
- 327 accounts targeted across eight organizations;
- 48 accounts hit with more than 1,000 failed attempts each.
What stands out most is how fast it worked once it got a foothold. In several cases, the time between first contact and privileged access to an organization's cloud infrastructure was under three hours — a pace a human operator would need days to match. Microsoft stresses the flaw wasn't in its own products but in a third-party supply chain. The company has already shipped detections for Defender XDR and Defender for Cloud Apps and sent containment guidance to affected customers.
The ARTEX episode is an uncomfortable preview of where AI-powered offensive security is heading. A tool meant to help defenders find holes before attackers do just showed what happens once it ends up in the wrong hands — closing the source code didn't stop it from leaking into exactly the use case its makers tried to prevent.
