Unit 42 says China-linked attackers ran the Cairn agent framework against 19 organisations
Twenty-five autonomous operations in August and September breached 10 targets, mostly semiconductor, defence-policy and civil-society organisations; Unit 42 says neither model refused the attackers' claim of an authorised security test.
- Security & misuse
- Notable
Palo Alto Networks’ Unit 42 reported that Chinese-speaking attackers used Cairn, an open-source AI penetration-testing framework, to run 25 operations against at least 19 organisations in August and September 2026. Cairn had already been named in Reuters-reported intrusions and in an Indian CERT advisory; this is the fullest public account of its use so far.
Unit 42 tracks the cluster as CL-STA-1238 and assesses “with moderate confidence” that it is China-nexus intelligence collection, citing Chinese-language tasking and authorisation files, exclusive use of Chinese models and targets that fit Chinese state priorities. In its description, the operators gave Cairn a target and a goal and AI “agent workers” planned and executed the steps. The attackers used a customised fork, with the models being Zhipu’s GLM 5.2 in 13 operations and DeepSeek’s V4 Flash and V4 Pro in 8 and 4. Operations ran from 20 minutes to more than 50 hours.
Targets included seven semiconductor and advanced-manufacturing organisations, three in defence and defence policy, and four in human rights and civil society. Unit 42 confirmed unauthorised access or data retrieval at 10. Its examples include more than a million member records at a US teachers’ union and roughly 150,000 internal documents at an industrial-automation manufacturer. The report says the weaknesses exploited were known ones: default credentials, authentication bypasses and dated public CVEs.
On guardrails, Unit 42 said Cairn’s stock prompts already tell the model never to refuse. The attackers added a Chinese-language preamble asserting an authorised security test and generated an authorisation file for each operation. Unit 42 wrote that “neither Zhipu GLM nor DeepSeek refused this framing”. It judged the group likely a single operator or small team, running 19 simultaneous intrusion attempts in about two weeks. Attribution rests on the vendor’s own telemetry and has not been independently confirmed.
A day earlier Unit 42 published a survey of 14 offensive-AI tools that counted 609 internet-facing ARTEX hosts and 1,323 running CyberStrikeAI as of October, and said ARTEX’s author had closed its repository “to prevent misuse”. It noted that other tools were filling the gap.
In the commentary
What people were saying around this time — external links, from the record's commentary rail.