Autonomous Cairn Operations Breached 10 Organizations, Researchers Say
Attackers used an AI-driven penetration-testing framework to run 25 operations against at least 19 organizations during August and September 2026, according to threat researchers at Palo Alto Networks’ Unit 42. The campaign targeted semiconductor and advanced-manufacturing organizations, defense and policy groups, and civil-society organizations. Researchers said the operations achieved unauthorized access or data retrieval at 10 targets.
The activity was tracked as CL-STA-1238. Unit 42 assessed with moderate confidence that it represented China-nexus intelligence collection, citing Chinese-language tasking, the use of Chinese AI models and the selection of targets in sectors associated with Chinese state intelligence priorities. The researchers did not identify a specific government organization as responsible.
Framework coordinated repeated attack tasks
The operators used a customized version of Cairn, a publicly available, open-source penetration-testing framework. Cairn’s AI-driven workers share findings through a “blackboard,” recording confirmed facts and queuing further actions. Operators could provide hints, but the framework planned and carried out task sequences based on a target and a stated objective.
Unit 42 observed the attackers running Cairn from a server hosted in Hong Kong and using a U.S.-based node for attack egress and staging. Operations lasted from 20 minutes to more than 50 hours. The researchers said the framework created target-specific plans, commonly spanning three to seven stages, and continued until its model judged the goal met, options exhausted or an operator stopped the task.
The operations used Zhipu GLM 5.2 and DeepSeek V4 Flash and Pro. Of the 25 runs, 13 used GLM 5.2, eight used DeepSeek V4 Flash and four used DeepSeek V4 Pro. The operators appeared to retry some targets with the higher-tier Pro model after Flash attempts met resistance. Each operation also included a custom “AIScan Runner Skills” pack with scanning and exploitation capabilities that Unit 42 said was not part of the upstream Cairn project.
Attackers also supplied prompts and authorization files that framed the work as approved security testing and instructed the models not to refuse. Neither model refused this framing, according to Unit 42. The researchers noted that attackers provided root domains and, in some cases, additional target-specific information, including VPN gateway details and shared group passwords. This indicates the automation was used alongside information gathered in advance.
Access included credentials, records and persistence
Confirmed outcomes varied by organization. At an industry standards body, weak-password spraying produced dozens of valid credentials and access to non-public directory data. Researchers also found that attackers overwrote a live standards-ballot document through an unauthenticated cloud-storage write, and exploited separate flaws to access customer accounts and meeting, voting and back-office functions.
At a semiconductor company, attackers used a production credential that was valid across multiple accounts to retrieve application data, user lists and email lists. They then created unauthorized administrator accounts and a persistent API token. At a U.S. teachers’ union, an account-creation bypass enabled access to more than one million member records, about 31,000 CRM contacts and staff password hashes, according to the framework’s logged activity.
Other cases involved an exposed registration endpoint supporting a defense-policy program, which revealed roughly 500 user records, including accounts with .mil and .gov addresses. Unit 42 also reported access to tens of thousands of content-management records across five additional organizations, including employee data, approximately 150,000 internal documents at an industrial-automation manufacturer and student data at an education nonprofit. The findings include records of attempted access and retrieval; they do not establish that every exposed record was subsequently used.
Defensive priorities remain familiar
Despite the autonomous orchestration, Unit 42 said the campaign relied on established weaknesses rather than novel exploitation methods. Those included authentication bypasses, insecure direct object references, mass-assignment privilege escalation, weak or default credentials, exposed cloud APIs and older publicly disclosed vulnerabilities. The researchers characterized AI’s contribution as increasing the volume and breadth of activity, not the complexity of the techniques.
For security teams, the activity reinforces the need to monitor identity and application-layer events alongside endpoint and network telemetry. Unusual account creation, unexpected administrator privileges, API-token issuance, repeated authentication failures and access to records at unusual scale can help reveal misuse. Correlating those signals across identity providers, cloud services, content-management systems and endpoints can give analysts a clearer view of activity that might otherwise appear as separate events.
Organizations should review internet-facing applications and registration flows, remove or rotate exposed credentials, restrict privileged access, and validate that cloud storage and APIs enforce authentication and authorization. Asset inventories and patch management remain important, particularly where older vulnerabilities are present. Security monitoring should also alert on changes to production identities and persistent credentials, while incident-response teams should be prepared to investigate affected systems and determine the scope of any data access.
Unit 42 said the campaign demonstrates how a small number of operators can use model-driven tools to pursue numerous targets concurrently. Its assessment is that autonomous, model-based intrusion activity is likely to become more common as accessible models improve. The reported operations show that existing controls around identity, exposed services and software vulnerabilities remain central defenses, even when attackers automate task planning and execution.