CrowdStrike says China-linked hackers used DeepSeek AI

A China-linked espionage group used an AI agent framework running mostly on DeepSeek's large language model to target more than 200 organizations around the world over 11 days, cybersecurity firm CrowdStrike ↘ 4.10% said in a report published Thursday, Oct. 29. NBC News Comcast Corporation ↗ 4.27% described the campaign as an example of how AI is letting hackers work at a scale that once took far more human effort.nbcnews.com+1

CrowdStrike tracks the group as VAPOR PANDA. The campaign started on Oct. 17 and was run through a framework called ARTEX, which reached entities in at least 14 countries from a single controller. The company confirmed post-exploitation activity at nine of the targeted organizations, according to CrowdStrike's blog post.crowdstrike.com

Who was targeted

The targets included foreign ministries in Latin America, Europe, Asia and the Middle East, as well as U.S. semiconductor, defense and think tank organizations and telecommunications providers in several regions, CrowdStrike said. In one engagement, the group went after a single European country's foreign ministry, presidency, prime minister's office and its defense, justice, finance and intelligence bodies. In another, it targeted 19 U.S.-based organizations in the pharmaceutical, semiconductor, defense, nuclear and financial sectors. The company said telecom providers in the U.S., Belgium and the Philippines were compromised. At least 12 more organizations, including U.S. think tanks and European government bodies, faced password spraying and credential scanning.crowdstrike.com

On compromised networks, the hackers installed BLOODALCHEMY, a backdoor that was first publicly documented in October 2023.crowdstrike.com

How the AI agents worked

According to CrowdStrike, ARTEX runs a main agent along with planner and worker agents that share a graph of facts and goals. Together they move each intrusion from initial access through reconnaissance and lateral movement. DeepSeek was the primary model, backed by a second subscription-based commercial model that CrowdStrike could not identify. In a single intrusion, the planning agents consumed about 1.4 billion input tokens.crowdstrike.com

The operation was not fully autonomous. Recovered transcripts showed a human operator giving instructions in Chinese, labeled "human instruction" and timestamped in China Standard Time. The operator stepped in at decision points and also told the agents to find hosts that lacked CrowdStrike's Falcon sensor, the company said.crowdstrike.com

An escalation in AI-driven espionage

CrowdStrike said this is VAPOR PANDA's first observed use of AI-driven intrusion tooling. The group has targeted government, think tank and energy organizations since 2021, and the campaign also marks a move into new sectors and regions. The company called the operation "an escalation in China-nexus use of AI agents, from single intrusions to multiple simultaneous engagements directed by a human operator." It said the campaign is separate from ARTEX activity other industry researchers have reported under the Storm-0940 name. In CrowdStrike's view, this likely means the framework is spreading among several China-linked groups, and relying on multiple AI model providers reflects a deliberate effort to build a resilient toolchain.crowdstrike.com

CrowdStrike's assessment is that agentic tools like ARTEX let "a small team to pursue a wider target set than manual operations would support."crowdstrike.com