SafetyTech
CrowdStrike attributes 200+ parallel agentic intrusions over 11 days to China-nexus VAPOR PANDA on DeepSeek + ARTEX.
The Hacker News · SecurityBrief · Unite.AI · 2026-10-30 · Source ↗
CrowdStrike attributes 200+ parallel agentic intrusions over 11 days to China-nexus VAPOR PANDA on DeepSeek + ARTEX.
The Hacker News · SecurityBrief · Unite.AI · 2026-10-30 · Source ↗
CrowdStrike's attribution report, picked up by The Hacker News, SecurityBrief and Unite.AI, documents the clearest run yet of multi-agent intrusions directed by a human operator. The vendor's threat intelligence team describes a China-nexus adversary it tracks as VAPOR PANDA, first identified February 2021 and overlapping with Trend Micro's Earth Estries cluster. Beginning October 17, the group ran the open-source ARTEX agentic framework across more than two hundred organizations in at least fourteen countries over an eleven-day window, drawing on the Chinese-developed DeepSeek model as the primary planner with a second commercial LLM in a backup role.
What ARTEX does, and how it was steered
ARTEX coordinates a main agent, planner agents and worker agents across shared intrusion graphs. The human operator issues instructions in Chinese at decision points and lets the agents handle enumeration, exploitation and lateral movement. CrowdStrike's accounting of one intrusion reports roughly 1.4 billion input tokens, almost entirely on DeepSeek. Post-exploitation activity was confirmed at nine organizations and included deployment of BLOODALCHEMY, a derivative of the ShadowPad implant, followed by credential spraying and network scanning. Reported targets include foreign ministries and organizations across pharmaceutical, semiconductor, nuclear and defense sectors.
"An escalation in China-nexus use of AI agents, from single intrusions to multiple simultaneous engagements directed by a human operator." — CrowdStrike attribution report, 2026-10-30
ARTEX is spreading across China-nexus groups
Microsoft linked ARTEX to a separate actor, Storm-0940, on October 21 via a ConnectWise supply-chain compromise. CrowdStrike says the VAPOR PANDA campaign is distinct from that activity and treats the overlap as evidence the open-source framework is being adopted across several China-linked groups. The report pairs with Epoch AI's October 29 finding, published the same week, that open-weight GLM-5.3 can autonomously compromise six recent Firefox releases at roughly $1,000 of compute per version, extending the attacker-defender ratio Anthropic and OpenAI flagged in their Q3 operational reports. The ARTEX attribution lands on the same day as OpenAI's post describing the four-layer monitoring stack that would have paged security thirty hours before the Hugging Face breach, framing the two publications as the first paired evidence of both sides of the agentic-attack-and-defense gap.