IncidentAI-enabled attacks
Unit 42: a China-nexus operator ran a fork of the Cairn AI pentest framework against 19 semiconductor, defense and civil society organizations
Palo Alto Networks Unit 42 reported on November 12, 2026 that a cluster it tracks as CL-STA-1238 ran 25 operations against at least 19 organizations from August to September 2026 with a customized fork of Cairn, an open-source AI penetration testing framework in which an AI model drives each agent worker and the workers coordinate through a shared board. Unit 42 confirmed unauthorized access or data retrieval at 10 of them. Targets were in the semiconductor and advanced manufacturing supply chain (7 organizations), defense and defense policy (3) and human rights and civil society (4), among them a Canadian human rights body. Examples include more than one million records from a US teachers' union, about 500 user records from a defense policy think tank or defense analytics vendor, persistent admin accounts at a semiconductor company and a live standards ballot document that was overwritten. The operator gave each task a one-line goal in Chinese and a root domain; Cairn mapped the target, split the goal into three to seven stages and looped until it was met, exhausted or stopped, in runs of 20 minutes to more than 50 hours. The models were Zhipu GLM 5.2 (13 operations), DeepSeek V4 Flash (8) and DeepSeek V4 Pro (4), with a skills pack that is not in upstream Cairn, and a Chinese-language authorization preamble and a generated AUTHORIZATION.md got past their guardrails. Unit 42 says the actors were likely a single operator or a small team.
Replay the attack step by step in the Learning Center
How it happened
- The operator gave Cairn agents a one-line goal and a target domain, and the agents mapped the target and worked through stages on their own.
- A fake authorization text and file kept the Chinese models from refusing.
- The agents used known weaknesses: authentication bypasses, IDOR, mass assignment, weak or default credentials, exposed cloud APIs and old public CVEs.
Impact
Unauthorized access or data retrieval at 10 of at least 19 organizations, including more than one million records from a US teachers' union and persistent admin accounts at a semiconductor company.
The fix
Unit 42 recommends identity hygiene, reducing the internet-facing attack surface and patching, since every weakness used was already known, and lists two IP addresses as indicators: 154.44.31[.]4 and 82.29.129[.]54.
Lessons
- Open-source agent frameworks put long, patient campaigns within reach of a single operator.
- Agents win with old, ordinary flaws; default credentials and unpatched CVEs on the edge are the first thing to fix.
- A model's guardrails are not a control when the attacker writes the authorization text.