AI Agents Enhance Chinese Espionage Tactics in Global Cyberattacks
A China-nexus espionage group has recently leveraged AI agents to conduct cyberattacks on over 200 organizations in just 11 days, as reported by CrowdStrike. The group, identified as Vapor Panda, began utilizing an agentic framework known as ARTEX on October 17, which enabled them to execute attacks across at least 14 countries. The targets included foreign ministries in Latin America, Europe, Asia, and the Middle East, as well as U.S. defense and semiconductor companies and telecom providers.
CrowdStrike confirmed that post-exploitation activities were observed at nine organizations, where the attackers installed the BloodAlchemy backdoor. Additionally, they employed password spraying techniques against at least 12 other entities, including U.S. think tanks and European government bodies. The ARTEX framework primarily operated on DeepSeek, alongside another unidentified commercial large language model (LLM). A main agent coordinated tasks among planner and worker agents, which maintained a shared graph of their learnings.
In one notable intrusion, planning agents processed approximately 1.4 billion input tokens. Transcripts revealed that operators intervened at critical decision points, issuing instructions in Chinese. When one attack vector failed, the agents quickly shifted to alternative methods, demonstrating a high level of adaptability.
For instance, at a pharmaceutical company, the AI agent exploited an unsigned SAML assertion in conjunction with an XXE vulnerability to access the cloud instance metadata service, ultimately escalating privileges to an enterprise admin account. In another case involving a Greek government entity, the operator provided the agent with a 2022 directory export from the organization’s mail server, which contained outdated authentication keys. The agent utilized these keys to forge an administrator token, gaining access to mailboxes and leveraging VPN configurations from a user’s sent emails to infiltrate the internal network.
This marks the first instance of Vapor Panda employing AI tools in their operations, and CrowdStrike notes that the group’s target list has expanded significantly as a result. The findings align with a separate report from Microsoft regarding ARTEX activity by another group, Storm-0940, suggesting a trend of increasing adoption of the ARTEX framework among China-linked cyber actors.
Kaspersky also detailed a smaller campaign targeting educational institutions in Africa and Latin America, where attackers utilized CyberStrikeAI, an open-source penetration testing framework. This framework operates on the attacker’s command-and-control server, generating commands for the implant on the victim’s machine, which executes the received instructions without any AI capabilities.
Detailed Tactics of Recent Cyber Intrusions
The attackers initially gained access through a compromised SQL Server, conducting reconnaissance for about a day before going silent for two weeks. Once the agent was deployed, it quickly established persistence and conducted local reconnaissance within two minutes, followed by domain mapping that took approximately five hours. Subsequent lateral movement and credential dumping across multiple hosts, along with a database export, were completed in less than an hour. The entire process from agent deployment to the compromise of domain admin accounts spanned three to four days.
While Kaspersky did not attribute this activity to a specific group, several indicators suggest involvement of Chinese-speaking operators, including Chinese comments in scripts and the use of a Chinese service for command-and-control operations. The development of CyberStrikeAI within the Chinese cybersecurity community further supports this hypothesis.
Historically, the use of AI by attackers has been limited to preparatory tasks such as crafting phishing emails and researching vulnerabilities. However, in these recent campaigns, AI has played a direct role in the intrusion process itself. Kaspersky researchers noted that the integration of AI into the attack chain does not introduce fundamentally new challenges for detection, as the actions taken on compromised hosts remain consistent with established attack methodologies.
Filed & Cleared
CrowdStrike emphasizes that detection measures are only effective where they are implemented, and these operators actively sought out vulnerabilities. They recommend organizations rotate keys and credentials following any suspected breaches, enforce SAML signature validation, and patch internet-facing devices to mitigate risks. The report concludes that China-nexus adversaries are likely to continue adopting agentic frameworks like ARTEX in the coming year, enhancing their capabilities to conduct large-scale intrusions efficiently.
Kaspersky researchers anticipate that as AI models evolve, the techniques employed by these attackers will become increasingly sophisticated, shifting the operator's role from task assignment to monitoring the AI's actions.




