Unit 42: Chinese-Speaking Actor Used AI Framework to Automate Attacks on U.S. Firms

Summary
Unit 42 reported that a Chinese-speaking actor used the Cairn AI penetration-testing framework with GLM and DeepSeek models to target at least 19 U.S. organizations, gaining access or exfiltrating data at 10.
Key points
- Unit 42 said the two-week operation in August and September targeted at least 19 U.S. semiconductor, defense and human rights organizations; it reported unauthorized access or data exfiltration at 10.
- The attackers configured Cairn with GLM and DeepSeek models, assigning different tasks based on model performance.
- Across 25 operations, Unit 42 reported 13 used GLM 5.2, eight used DeepSeek V4 Flash and four used DeepSeek V4 Pro.
- A custom skill pack added scanning and exploitation capabilities not present in the upstream Cairn project.
- The agents coordinated through a shared system, and operators could steer them. Unit 42 said Chinese-language instructions framed tasks as authorized security work.
- Unit 42 described one target's attack chain as reconnaissance, web application and middleware exploitation, credential exfiltration, privilege escalation and internal-network persistence.
- Unit 42 said the attackers took credentials from an industry standards body, one million teachers' union member records, 500 records from a defense-policy think tank and 150,000 internal documents from an industrial-automation manufacturer.
Article Details
- Attack Vectors
- The operator configured the Cairn AI penetration-testing framework to use Zhipu GLM and DeepSeek models, using a Chinese-language pretext that framed tasks as authorized security testing to bypass the models’ limited safeguards.
- Cairn agents used a campaign-specific skill pack with scanning and exploitation capabilities to develop target-specific attack plans.
- In an attack against a semiconductor manufacturer, the reported chain included reconnaissance, web application and middleware exploitation, credential exfiltration, privilege escalation, and persistence on the internal network.
MITRE ATT&CK
T1190 · Exploit Public-Facing ApplicationIn the reported attack against a semiconductor manufacturer, the chain included web application exploitation.T1595 · Active ScanningThe reported attack chain began with reconnaissance, and the campaign-specific Cairn skill pack provided scanning capabilities.
Products
DeepSeek V4 FlashOut of 25 operations, the attackers ran 13 of them on GLM 5.2, eight on DeepSeek V4 Flash and four on DeepSeek V4 Pro, "using the higher-tier DeepSeek V4 Pro for repeated attempts against targets that resisted theDeepSeek V4 ProOut of 25 operations, the attackers ran 13 of them on GLM 5.2, eight on DeepSeek V4 Flash and four on DeepSeek V4 Pro, "using the higher-tier DeepSeek V4 Pro for repeated attempts against targets that resisted theGLM-5.2Out of 25 operations, the attackers ran 13 of them on GLM 5.2, eight on DeepSeek V4 Flash and four on DeepSeek V4 Pro, "using the higher-tier DeepSeek V4 Pro for repeated attempts against targets that resisted theZhipu GLMThe attackers configured the Cairn framework to run on Chinese AI models Zhipu GLM and DeepSeek. For two weeks in August and September, the operation targeted at least 19 organizations, successfully gaining unauthorized
Tools
Countries
Japancyberattacks. The same method had been used to breach numerous South Korean financial institutions and Japanese companies in recent weeks.South Koreaof faster, cheaper and more accessible cyberattacks. The same method had been used to breach numerous South Korean financial institutions and Japanese companies in recent weeks.United Statesused an open-source artificial intelligence penetration testing framework to automate attacks against U.S. semiconductor manufacturing, defense and human rights organizations, researchers found.
Industries
Defenseintelligence penetration testing framework to automate attacks against U.S. semiconductor manufacturing, defense and human rights organizations, researchers found.Human rights organizationstesting framework to automate attacks against U.S. semiconductor manufacturing, defense and human rights organizations, researchers found.semiconductor manufacturingartificial intelligence penetration testing framework to automate attacks against U.S. semiconductor manufacturing, defense and human rights organizations, researchers found.
