VAPOR PANDA used ARTEX agents in parallel intrusions; nine organizations had confirmed post-exploitation

CrowdStrike reports that a VAPOR PANDA operator ran ARTEX and language models, primarily DeepSeek, across multiple intrusion tasks from October 17 onward. The controller reached more than two hundred organizations as targets; CrowdStrike confirmed post-exploitation activity at nine, including deployment of BLOODALCHEMY and follow-on activity. Those counts must not be interchanged. A trace of 1.4 billion input tokens in one engagement describes model workload, not independent attacks or victims. The report distinguishes this campaign from separate ARTEX activity labeled Storm-0940 by other investigators.

Individual victim examples show different ordinary intrusion prerequisites beneath the agent framework. One pharmaceutical path combined an improperly accepted SAML assertion with XML external entity access to metadata and a cloud role; another involved an exposed orchestration service, root inside a container and Active Directory credential/delegation moves. A third government network case relied on historical mail authentication keys and a reportedly leaked VPN client key. Agentic planning coordinated exploration, but none of those observations establishes a novel exploit inherent to the model itself, a kernel escape from the container, or a verified compromise of every target on the list.

The meaningful AI component is the parallel planner and worker workflow using obtained credentials and tool feedback under human direction. Effective controls must bind each tool to the intended principal and target, while incident analysis preserves the original vulnerability and the separate evidence of subsequent AI-assisted activity.