Insights

AI Security

A customized Cairn fork ran 25 operations against at least 19 organizations

Unit 42 reports Chinese-speaking attackers used a customized fork of the open-source Cairn framework in 25 operations against at least 19 organizations, which it assesses with moderate confidence as China-nexus collection. Separately, a criminal crew chained Strix and Cairn against online retailers.

November 19, 2026 · 5 min read

A customized Cairn fork ran 25 operations against at least 19 organizations

Unit 42 reports that from August to September 2026, Chinese-speaking attackers used a customized fork of Cairn, an open-source AI penetration testing framework, to run 25 operations against at least 19 organizations, with unauthorized access or data retrieval at 10 of them. Unit 42 assesses with moderate confidence that this is China-nexus intelligence collection, and tracks it as CL-STA-1238. Moderate is the right word, and I am keeping it.

This is part of a short series on how threat actors are using agentic offensive tools right now. Cairn gets its own post because two very different kinds of operators picked it up within weeks of each other.

What CL-STA-1238 did with Cairn

Unit 42 describes Cairn as a framework in which AI-driven agent workers coordinate through a shared blackboard, recording confirmed findings as facts and queuing next steps as intents, while a human can add hints at any time. In this campaign, Unit 42 says the attackers supplied a target and a goal, and the framework planned, executed, and recorded results on its own.

The targets were focused. Unit 42 lists seven organizations across the semiconductor and advanced manufacturing supply chain, three in defense and defense policy, and four in human rights and civil society. Unit 42 reports that individual operations ran from 20 minutes to more than 50 hours with no time limit set by the attackers, and that confirmed outcomes included valid credentials, persistent administrator accounts in a production system, and exposure of more than one million member records at one organization.

The attribution rests on three things, according to Unit 42: Chinese-language tasking and artifacts, exclusive use of Chinese AI models (Zhipu GLM and DeepSeek), and a target set that matches established Chinese state intelligence priorities. That supports moderate confidence. It is not a named group, and I would not brief it internally as one.

GTIG separately reports observing Cairn in intrusion activity against targets including the US aerospace sector, without tying that activity to CL-STA-1238, so I do not tie it either.

The weaknesses were old ones

The part of the report I would hand to every IT leader is the list of what got exploited. Unit 42 lists authentication bypasses, insecure direct object references, mass-assignment privilege escalation, weak and default credentials, exposed cloud APIs, and dated public CVEs. Its conclusion is that autonomous AI raises the volume and coverage of exploitation, but not its complexity.

Unit 42 also assesses that the attackers were likely a single operator or small team, running simultaneous attempts against 19 organizations in two weeks. That is the shift worth planning for. The old limit on how many targets one crew could work at once was operator hours, and an agent working the list removes it.

A criminal crew chained Strix in front of Cairn

Cairn is not only an espionage tool. Gambit Security reported a financially motivated operator who ran three open-source AI harnesses against online retailers: Strix for vulnerability discovery, Cairn for autonomous exploitation, and Hermes to orchestrate the campaign. Gambit found that Strix ran 146 times in deep mode against 138 hosts between August 23 and 31, 2026, and that some of those reports became the opening of the Cairn stage.

Per Gambit, between September 10 and 15 alone, 105 attack projects were launched and at least 27 companies were compromised to varying degrees, with activity going back to July 2026. Gambit also reports that where access was achieved it usually took less than a day, and that data was destroyed at some victims by cleanup steps the agent ran.

Gambit's report shows this operator working in Chinese, but none of the reporting links the retail campaign to the Unit 42 cluster. Keep the two lanes separate: one moderate-confidence China-nexus collection campaign, one criminal retail campaign, same open-source engine.

Exposed consoles cut both ways

NorForge found more than 20 exposed Cairn panels it assessed as likely tied to malicious activity, plus a considerable number of apparently legitimate red team operations exposed the same way. For one actor cluster's successful projects, NorForge measured an average of a little under 23 hours from project start to goal.

Two lessons follow. Agent consoles are internet-facing infrastructure, so if anyone in your organization runs agentic tooling, including for authorized testing, the console needs authentication and a named owner. And NorForge's 23 hours is the planning number to hold next to the second-scale bursts in other reporting. Elastic makes the same point: individual bursts fit a short detection window, but operations run for hours, so connecting them takes a long lookback.

What I would do this quarter

Test the IR plan against volume, not sophistication. Run a tabletop where three ordinary weaknesses, such as a default credential on a partner portal, an open registration endpoint, and a writable cloud storage location, are found in the same afternoon by something that does not sleep. Whose queue does each one land in, and does anyone see that they are one operation?

Keep behavioral detections current on the boring things: password spraying against internet-facing apps, new administrator accounts and API tokens in production systems, account creation through endpoints that should be closed, and bulk record reads from one session. Unit 42's point is that identity hygiene, attack-surface reduction, and patching still work. Detections should tell you quickly when one of them has slipped.

Be able to respond without a meeting. When a new admin account appears in a production system outside change control, containment should not wait for someone to read the alert. Pre-approve the action, disable the account, revoke its tokens, and have a person review the decision afterward. Elastic's guidance is to decide which high-confidence detections start containment before an incident begins.

The question for your next tabletop: if an agent spent 50 hours working your external surface, which alerts would have fired, and who would have connected them?