More than two hundred targets, nine confirmed compromises. The China-linked group that CrowdStrike tracks as Vapor Panda used ARTEX, a system coordinating multiple artificial intelligence agents, against more than 200 organizations in at least 14 countries. Observed since October 17, this campaign combines automation of intrusions and human interventions at decision points.
Targets include administrations, telecom operators and companies in the pharmaceutical, financial, nuclear and semiconductor sectors. The report describes a cyberespionage campaign, without establishing a link with a crypto platform hack.
Key Points
Vapor Panda has targeted more than 200 organizations in at least 14 countries, according to CrowdStrike.
Researchers confirm post-compromise activities at nine organizations.
ARTEX distributes tasks between several agents, while a human operator maintains a management role.
These observations illustrate an increasing automation of attacks, without demonstrating that AI can carry out all operations alone.
ARTEX coordinates several intrusions under human supervision
ARTEX functions as a set of agents that can use tools to examine a network, search for vulnerabilities, and execute actions. A master agent coordinates planning and execution agents, who share their findings.
According to CrowdStrike, Vapor Panda primarily used DeepSeek, as well as a second, unidentified business model. The recovered conversations show an operator giving instructions in Chinese, in particular to obtain identifiers or install persistent access.
Automation therefore does not mean the absence of supervision. Humans guide the agents and intervene when important decisions arise. The researchers observed several simultaneous operations from the same ARTEX controller, which makes it impossible to determine how many people were participating in the campaign.
The precedent published by Anthropic in November 2025 has some similarities. The company described a campaign attributed to a group supported by the Chinese state, targeting around thirty organizations with Claude Code. She then estimated that the AI had executed 80 to 90% of tactical operations. However, these figures relate to another survey and do not directly measure ARTEX's progress.
For crypto, a risk that goes beyond smart contracts
The CrowdStrike report does not single out crypto platforms as a particular target. However, the techniques observed also concern companies that retain or transfer digital assets: compromise of accounts, theft of identifiers and exploitation of exposed services.
The Bybit hack in February 2025, for around $1.46 billion, is a reminder of this exposure. Investigations identified malicious code in the Safe{Wallet} interface, after a developer was compromised. The attack was not based on a classic flaw in a DeFi protocol. However, no element cited here allows this operation to be attributed to ARTEX or to AI agents.
The attack targeting the npm software catalog in September 2025 also illustrates the risks associated with code providers. A phishing attack allowed eighteen popular JavaScript libraries to be modified to introduce code targeting crypto transactions. The compromise of a shared component can therefore expose numerous applications.
