A China-nexus hacking group used an automated, agentic artificial intelligence penetration-testing tool to target more than 200 organisations in at least 14 countries in 11 days from 17 October, according to cybersecurity firm CrowdStrike. The group, which CrowdStrike tracks as Vapor Panda, paired the open-source ARTEX framework with the DeepSeek model to run multiple attack paths at once, at a pace CrowdStrike says reached more than 1,500 commands an hour, against a historical hands-on-keyboard norm of up to 40.

CrowdStrike told CyberScoop that about 70% of the targets, roughly 140 organisations, were confirmed compromised or showed moderate-confidence evidence of intrusion. That figure needs care: CrowdStrike's own blog confirms post-exploitation activity at nine organisations, where the group deployed the BLOODALCHEMY backdoor, sprayed credentials and scanned networks. The wider 140 figure is CrowdStrike's assessment as relayed by CyberScoop, not a list of named victims, which the sources do not publish.

Who was targeted

Targets included foreign ministries, a European country's presidency and prime minister's office, defence and intelligence agencies, and telecom providers in the United States, Belgium and the Philippines, along with 19 US-based organisations in pharmaceuticals, semiconductors, defence, nuclear and finance. CrowdStrike says the campaign remains active and is separate from other ARTEX activity that Microsoft attributed to a different group, Storm-0940.

A wider shift

CrowdStrike reported observing roughly 45 agentic adversary attacks in September and 126 in October at the time of reporting, suggesting AI-assisted intrusion is becoming routine rather than exceptional. Attribution to Vapor Panda is CrowdStrike's assessment, and the group is described as China state-backed; the firm did not identify a second commercial model used alongside DeepSeek. For defenders, the practical point is speed: an attack that once took a human team days can now probe many targets in parallel, which makes patching, logging and credential hygiene harder to defer.