One operator, two hundred targets
CrowdStrike says a China-linked group it calls Vapor Panda used ARTEX, an AI agent framework running on DeepSeek and a second commercial model, to work against more than 200 organisations in 11 days. Nine were confirmed compromised.
Poa Mosyuen, CC BY-SA 3.0, via Wikimedia CommonsA China-linked hacking group used an AI agent framework called ARTEX, running mainly on DeepSeek's model, to target more than 200 organisations in at least 14 countries over 11 days from 17 October, CrowdStrike said in a report on Thursday. The security company confirmed hands-on intrusions at nine of them.
CrowdStrike calls the group Vapor Panda and describes it as a "China-nexus" adversary active since 2021, mostly against governments, think tanks and energy and extractive companies. ARTEX, the report says, runs a main agent that directs planner and worker agents through a shared map of facts and goals, taking each target from first access through reconnaissance to moving across the network. The second model was a subscription commercial one that CrowdStrike "could not confirm"; it does not name it. In a single intrusion, the planning agents consumed about 1.4bn input tokens.
The human role was small and visible. CrowdStrike says it recovered transcripts of an operator giving the agents instructions in Chinese, labelled "human instruction" and time-stamped in China Standard Time, asking for web shells, domain administrator passwords and an inventory of security software. At one pharmaceutical company, the operator told the agents to find Windows machines "without Falcon", CrowdStrike's own endpoint sensor. Adam Meyers, CrowdStrike's vice president of threat intelligence, told NBC News, as Fox News reported: "This is the first time we've seen them using an AI or agentic capability at scale."
| Victim type, as CrowdStrike describes it | What CrowdStrike says happened |
|---|---|
| Foreign ministries of two Latin American countries | Extensive intrusion; BLOODALCHEMY backdoor at one |
| An interior ministry in the Middle East | Compromised; BLOODALCHEMY deployed |
| Telecoms providers in the US, Belgium and the Philippines | Compromised |
| Two US pharmaceutical organisations | Compromised, from an engagement aimed at 19 US organisations |
| A Greek government entity | Mailboxes read using unrotated 2022 keys and a VPN key from a 2024 email leak |
| US think tanks and European government bodies | Password spraying and credential scanning, at least 12 organisations |
The techniques were not new; the throughput was. Agents chained an unsigned SAML login assertion with an XML flaw to reach a cloud metadata service and take over a role that could command thousands of servers, according to the report. Against the Greek target, the operator supplied directory-export keys from 2022 that had probably never been rotated, and the agents used a VPN key from a May 2024 email leak to get inside. The agents tried several paths at once and switched when one failed.
By our arithmetic, more than 200 organisations in an 11-day window is about 18 a day, and nine confirmed intrusions is fewer than one in twenty of the targets. Adding the 12 organisations hit with password spraying gives 21 with confirmed hostile activity, about one in ten. AI lets such groups "operate on a scale that previously required far more human effort", NBC News reported. Neither CrowdStrike nor NBC carried a response from DeepSeek.
This is the second ARTEX campaign made public in little more than a week. Microsoft said last week that a different China state-aligned group, Storm-0940, used the same framework after compromising a backup vendor to try sign-ins at more than 2,250 customer tenants, as Komposite reported on 22 October. CrowdStrike says the Vapor Panda activity is separate, and that it has also seen ARTEX used by criminal and unattributed actors.
Why it matters: an AI agent that holds credentials behaves like any other privileged account, a point Komposite made in September: the agent is a privileged user. What changes is the arithmetic for defenders, since one operator can now run many intrusions in parallel. CrowdStrike assesses that China-nexus groups will "highly likely" keep adopting such frameworks over the next 12 months, and its first recommendation is unglamorous: rotate signing keys and service credentials, including those sitting in old backups.