Category: AI Agent Security | Reading time: 3 min
ARTEX, the AI offensive security framework that was misused against South Korean financial institutions, keeps turning up in real intrusions. Incident-response firm S-RM published what it saw while responding to several severe October intrusions at separate organizations, and blockchain infrastructure provider Chainstack disclosed a breach that reporting ties to the same campaign.
What S-RM saw
According to S-RM, the attacks began by probing internet-facing infrastructure and exploiting a web server through a vulnerable plugin or a misconfigured upload. From there the agents pulled credentials, tried to break out to the management plane, and repeated a loop: find a credential, test it everywhere, move on.
In one intrusion, the attacker went from initial access to all identity infrastructure, cloud, source code and perimeter firewalls in 24 hours. In another, it went from a web server to central management platforms in four hours, running two to three attack chains at once. When responders blocked an IP address, the activity pivoted within 10 minutes to Tor exit nodes and more than 50 new addresses.
S-RM notes that ARTEX was made private after the Korean misuse, but forks remain available. In several cases the attacker's intent could not be confirmed because the intrusion was interrupted mid-movement. S-RM expects more incidents.
Chainstack
Chainstack says an attacker had access from October 26 to 31 through a third-party component, reaching its development environment and parts of production. Exposed data included organization names, account holder names, member emails and password hashes. Node credentials sat in a reachable system; Chainstack says it has no evidence they were read but is treating them as exposed and has told customers to rotate node credentials, regenerate API keys and update passwords.
AI Weekly, citing The Hacker News, links the breach to the ARTEX campaign also seen at WorkOS, Chainalysis and DIVD. Attribution is not settled: SecurityWeek links ARTEX use to Storm-0940, a group Microsoft tracks as Chinese, while AI Weekly cites CrowdStrike naming a cluster "Vapor Panda."
Why it matters
A 24-hour path from one web server to every identity system leaves little room for review after the fact. The organizations that can answer "what did this identity do, and when" are the ones that recorded each action as it happened. The same applies to the agents enterprises run themselves: scoped, revocable credentials per agent and timestamped records of every action are what let a team show outsiders what happened.
Source: S-RM, "ARTEX in the wild", AI Weekly on Chainstack, SecurityWeek