IncidentAI-enabled attacks
VAPOR PANDA: ARTEX agents on DeepSeek went after more than 200 organizations in 14 countries in 11 days
CrowdStrike Counter Adversary Operations reported on October 29, 2026, that VAPOR PANDA, a China-nexus targeted-intrusion adversary that has hit government, think tank, energy and extractive organizations since 2021, made its first observed use of AI-driven intrusion tooling. From October 17 a single controller of ARTEX, a framework that coordinates a main agent with planner and worker agents, directed activity against more than 200 organizations in at least 14 countries over an 11-day window, mainly with the DeepSeek model and a second subscription model CrowdStrike could not confirm; the operator stepped in only to steer the agents at decision points. Targets included foreign ministries in Latin America, Europe, Asia and the Middle East, presidency and prime minister offices, defense, justice, finance and intelligence bodies, US think tanks, and 19 US organizations in pharmaceuticals, semiconductors, defense, nuclear and finance. Telecom providers in the US, Belgium and the Philippines were compromised. CrowdStrike confirmed post-exploitation at nine organizations and password spraying or credential scanning at at least 12 more; it told CyberScoop that about 70 percent of targets were confirmed compromised or showed evidence of a likely intrusion. In one intrusion the planning agents consumed about 1.4 billion input tokens. CrowdStrike says the campaign is separate from Microsoft's Storm-0940 ARTEX activity.
Replay the attack step by step in the Learning Center
How it happened
- Unsigned SAML assertions combined with an XML external entity flaw to query the cloud metadata service, plus an unauthenticated orchestration service that gave root in a production container environment.
- Old leftovers turned into access: stale keys from a 2022 mail server export forged an administrator token, and VPN settings and a client key were recovered from mail and an earlier leak.
- Agents looked for Windows machines without the Falcon sensor, then the BLOODALCHEMY backdoor, LAPS password reads, delegation abuse, mailbox reads and Chisel and SoftEther tunnels followed.
Impact
Confirmed post-exploitation at nine organizations, compromised telecom providers in three countries and ministries in several regions; CrowdStrike's Adam Meyers told CyberScoop the activity could exceed 1,500 commands an hour, against about 40 before.
The fix
CrowdStrike advises rotating signing keys, preauthentication keys and service credentials after any suspected data exposure, enforcing SAML signature validation, requiring IMDSv2 and restricting cloud role permissions, patching internet-facing appliances and long-lived Linux servers, deploying endpoint detection on every server including those not reachable from the internet, and hunting for ARTEX markers such as artx or artex in command lines, artx-prefixed cloud roles, art18- comments and C:\ProgramData\artx5v\.
Lessons
- Agents make old exposures cheap to use at scale: a leaked key or a forgotten service is now found and tried across hundreds of targets at once.
- An attacker's agents go where detection is missing, so coverage gaps matter more than ever.
- Rotate credentials after any leak, even years later; this campaign used keys from 2022.