China-Nexus Group Uses AI Framework to Breach 10 Global Organisations
Share
A China-linked threat actor has successfully compromised at least 10 organisations, including United States defence contractors and semiconductor manufacturers, by deploying an autonomous artificial intelligence (AI) framework to automate its operations.
The campaign, identified by researchers at Palo Alto Networks Unit 42, utilised an open-source AI penetration-testing tool known as Cairn. The activity, tracked under the designation CL-STA-1238, targeted 19 organisations in total between August and September, achieving unauthorised access or data retrieval in more than half of the attempts.
Autonomous Execution and AI Models
The attackers leveraged Cairn to move away from manually directed intrusions. By providing the framework with a specific target and a high-level objective, the AI agents planned and executed reconnaissance, identified vulnerabilities, and attempted exploitation with minimal human intervention.
According to researchers, the threat actors weaponised the framework using Chinese large language models (LLMs), specifically Zhipu GLM 5.2 and DeepSeek V4 Flash and Pro. These models drive individual agent workers that coordinate their activities through a shared “blackboard” system, allowing the tools to build upon previous findings without constant human oversight.
The operations varied in duration, with some lasting only 20 minutes while others continued for more than 50 hours. This automation allowed the group to manage multiple simultaneous intrusions across different sectors.
Targeted Industries and Data Exposure
The campaign focused heavily on critical supply chains and sensitive policy sectors. Seven of the targeted entities were involved in semiconductor manufacturing, while three were linked to US defence and defence policy think tanks. Other victims included human rights organisations and a teachers’ union.
In one instance, the breach resulted in the exposure of more than 1 million personal records. In other successful attacks, the group obtained dozens of valid credentials, retrieved military and government identities, and created persistent administrator accounts within production systems. At one industry standards body, the attackers successfully modified a live standards-ballot document.
Traditional Vulnerabilities Remain the Entry Point
Despite the sophisticated use of AI for automation and lateral movement, the initial points of entry remained largely conventional. The threat actors exploited well-known security gaps that continue to plague many enterprise environments.
The researchers noted that the attackers utilised authentication bypasses, weak or default credentials, and dated public common vulnerabilities and exposures (CVEs). Brute-force attacks and password spraying were also employed against partner-integration applications and identity platforms.
Identity hygiene and robust patch management remain the primary defences against these automated threats. Security experts recommend that organisations prioritise attack-surface reduction and the implementation of cyber deception, such as tripwires, to detect autonomous agents during the reconnaissance phase before they achieve their primary objectives.




Leave a Reply