Multi-agent attack frameworks in the field: CyberStrikeAI and ARTEX

A CSA research brief summarizing multiple reports states that the open-source multi-agent penetration testing frameworks CyberStrikeAI and ARTEX have been used in real intrusions in Japan, South Korea and elsewhere, by both state-backed groups and individual criminals. Microsoft said the China state-backed Storm-0940 used stolen third-party application identities to sign in to more than 2,250 Microsoft 365 tenants; South Korean police said a domestic group used ARTEX in most of its attacks, breaching more than 90 organizations and stealing about 8.2 million records; TrendAI found CyberStrikeAI attacking at least 19 organizations with 44 agents. The entry points disclosed in the reports remain exposed management interfaces, default or reused credentials and over-privileged third-party application identities, with the frameworks mainly amplifying the scale and speed of attacks. The operator of the South Korean bank campaign ran ARTEX mainly on DeepSeek v4.1-flash, possibly accessed through an API reseller; because the framework accepts any OpenAI-compatible endpoint, a single provider's safety restrictions are difficult to reliably block.