
Palo Alto Networks' Unit 42 published research on 12 November 2026 about a campaign it tracks as CL-STA-1238. In August and September 2026, attackers ran 25 operations against at least 19 organisations and confirmed unauthorised access or data retrieval at 10 of them. Unit 42 assesses with moderate confidence that this is China-nexus intelligence collection, citing Chinese-language artifacts, the use of Chinese AI models and the choice of targets. It believes a single operator or a small team was behind it.
How it worked: the attackers used a customised fork of Cairn, a public open-source AI penetration-testing framework. An AI model drives each "agent worker", and the agents share findings and queued actions on a common "blackboard". Operators supplied a target and a one-line goal; Cairn mapped the target, split the goal into small tasks and kept going until it judged the goal met. Runs lasted from 20 minutes to more than 50 hours. Of the 25 operations, 13 ran on Zhipu GLM 5.2, 8 on DeepSeek V4 Flash and 4 on DeepSeek V4 Pro; the attackers appear to have switched to V4 Pro for targets that resisted Flash.
No guardrail stopped it. Cairn's stock prompts tell the model never to refuse a task, and each operation added a Chinese-language claim of authorised testing and a generated AUTHORIZATION.md file. According to Unit 42, neither model refused.
What they got: more than 1 million member records and about 31,000 CRM contacts at a US teachers' union through a CMS flaw; about 500 user records, including .mil and .gov accounts, through an open registration endpoint on a defence-policy platform; roughly 150,000 internal documents at an industrial-automation maker; persistent administrator accounts and an API token in a semiconductor company's production system; and dozens of valid credentials at an industry standards body, where a live standards-ballot document was overwritten. Targets included seven semiconductor and advanced-manufacturing firms, three defence-related organisations and four human-rights and civil-society groups in the US and Canada.
No new zero-days were needed. The weaknesses were authentication bypasses, insecure direct object references, mass-assignment privilege escalation, weak and default passwords, exposed cloud APIs and old public CVEs. Unit 42's Andy Piazza told CSO Online the approach "completely breaks our attribution models" and that such a campaign "can now interact with 1,000 organizations at the same time" — a description of capability, not a count of victims. Zhipu and DeepSeek have not commented in the coverage we read.

