Chinese AI pentest agent ARTEX used in Japan breach wave hitting ~20 companies, GMO says

#What happened

GMO Cybersecurity by Ierae, the security arm of Japan's GMO Internet Group, confirms that ARTEX - an AI tool developed in China to test cyberdefences - was used in many of the recent attacks on Japanese companies, cybercrime countermeasures director Daiki Fukumori said in a blog post (reported by Bloomberg on Oct 20). The attacks were launched from a single server infrastructure aimed almost exclusively at Japan, targeting about 20 Japanese companies, including nine in the same industry. The findings land amid a breach wave affecting firms from Daiwa Securities Group to East Japan Railway Co and SoftBank. Ierae also found six more AI-enabled attack servers targeting Japanese companies across e-commerce, finance, retail, logistics, media, education and marketing. Eight of the targeted companies have disclosed data leaks in the last two months, and Ierae believes one group may be behind multiple breaches. It has shared its findings with law enforcement and is investigating whether countries beyond Japan and South Korea are also being targeted. The Japan campaign extends documented ARTEX use beyond South Korea: CrowdStrike reported on Oct 7 that the same tool was used in a campaign against South Korean banks from late September to early October. The Wall Street Journal reported Oct 21 that the investigation covered six overseas servers examined between Sept 18 and Oct 18, and that attribution to specific people remains inconclusive.

#Why it matters

Autonomous AI agents are moving from demos to active attack infrastructure. ARTEX automates the work of a human penetration tester - analyzing targets, planning intrusion routes, and operating security tools - which compresses the time between disclosure and exploitation for defenders. For enterprises, the practical takeaway is that version-based patching and firewalls are weakening as safety nets when AI agents can probe unlisted software versions and rotate IPs below rate-limit thresholds. And for the AI labs racing on agent capabilities, this is another datapoint in the rogue-agent debate: the same agentic techniques that power coding assistants are now standard tooling for intrusion campaigns. Verified Oct 22, 2026.