Autonomous intrusion agents reach state-aligned espionage: what Unit 42’s Cairn case shows
Unit 42 says a China-nexus cluster ran 25 largely autonomous intrusion operations against semiconductor, defence and civil-society organisations, with confirmed access at 10. The exploited weaknesses were ordinary; the speed and parallelism were not.
The finding
Agentic intrusion tooling is now part of state-aligned collection, not only criminal experimentation. On 12 November Palo Alto Networks Unit 42 described a cluster it tracks as CL-STA-1238, which ran a customised fork of the open-source Cairn framework against semiconductor, defence and civil-society targets [1]. Unit 42 assesses with moderate confidence that the activity is China-nexus intelligence collection. It bases that on Chinese-language tasking and authorisation files, exclusive use of Chinese models, and target selection it says aligns with Chinese state priorities [1].
The operators supplied a root domain and a goal. The framework then planned, executed and recorded its own steps, with humans able to inject hints [1]. One translated task name read “breach the target as fast as possible” [1].
What the data says
Unit 42 counts 25 operations against at least 19 organisations between August and September 2026, with simultaneous attempts against all 19 inside roughly two weeks [1]. Single operations ran from about 20 minutes to more than 50 hours, with no time limit set by the operators [1]. Unauthorised access or data retrieval was confirmed at 10 organisations [1].
- Semiconductor and advanced manufacturing: seven organisations across lithography, chip design, materials, photomask and standards. Unit 42 reports persistent administrator accounts and an API token created at a semiconductor company, and a live standards-ballot document overwritten through an unauthenticated cloud-storage write [1].
- Defence and policy: a space technology company, a defence-analytics vendor and a policy think tank. About 500 identity-platform user records, including .mil and .gov accounts, were retrieved from the policy target [1].
- Civil society: a Canadian human-rights body, a US teachers’ union, a science-policy group and an education non-profit. More than one million member records were taken from the union [1].
- Industrial automation: roughly 150,000 internal documents at a manufacturer [1].
The models were Zhipu GLM 5.2 in 13 operations, DeepSeek V4 Flash in eight and DeepSeek V4 Pro in four. Unit 42 notes an apparent move to the larger model against targets that resisted the smaller one [1]. A Chinese-language preamble instructed the model not to refuse, and each operation generated a file declaring the work client-approved. Neither model refused [1].
Cairn itself is a public AGPL-licensed project. Its README describes a blackboard architecture of facts, intents and hints, lists penetration testing and CTF solving as intended uses, and restricts use to authorised work [2]. Unit 42 says the operators added a proprietary skills pack for scanning and exploitation that is not in the upstream code [1].
Where this sits
CL-STA-1238 is the latest of several cases. In November 2025 Anthropic reported a group it designated GTG-1002, assessed with high confidence as Chinese state-sponsored, that used Claude Code to perform an estimated 80 to 90 percent of tactical work against about 30 targets. It also noted the model overstated findings, which forced operators to validate results [3]. In September 2026 Hunt.io described a Chinese-speaking operator using a multi-agent framework it calls SecFlow against government and education systems in Taiwan, Indonesia, mainland China and Vietnam [4].
The ARTEX framework has driven most recent reporting. Microsoft assesses that Storm-0940, a China state-aligned actor, used ARTEX in the ConnectWise Cloud Backup supply-chain compromise and attempted sign-ins to more than 2,250 tenants [5]. Microsoft separately reported a Cambodia-based financially motivated cluster, Storm-3358, using ARTEX against financial organisations mainly in Southeast Asia [6]. It also documented ARTEX moving from GitHub secret discovery to Key Vault exfiltration in five minutes in one case [7]. Arctic Wolf has linked AntiFox, an ARTEX fork run from Alibaba Cloud, to intrusions that read production AWS secrets after hundreds of API calls in about 30 minutes [8][9][13].
Estimates of how widely these consoles are deployed differ. Censys counted 1,465 exposed ARTEX hosts in early November, or about 815 once two identically configured hosting blocks are excluded [10]. Check Point counts 2,557 systems that host or have hosted an ARTEX panel and 1,608 live instances, while its own key points cite more than 1,800 [11]. The methods and dates differ, and the figures should not be read as one number. Censys also notes that some exposed hosts are likely run by authorised testers [10].
What it means
Unit 42’s central judgement is that autonomous AI raises the volume and coverage of exploitation, not its complexity [1]. Every weakness in its report was known: password spraying, credential reuse across partner integrations, open registration endpoints, authentication bypasses, insecure direct object references and mass-assignment escalation [1]. Check Point’s criminal campaigns show the same pattern, with agents exploiting hardcoded credentials and leaked storage keys [11].
The practical change is tempo and parallelism. One operator can now hold 19 targets open at once and let the agent grind for two days. That fits the UK NCSC’s 2025 judgement that AI will almost certainly make intrusion more efficient and widen a divide between systems that keep pace and those that do not [12]. It also means the long tail of weak authentication in smaller suppliers, standards bodies and non-profits becomes a cheap route into high-value supply chains.
For Malaysia and the wider region, two points stand out. Microsoft’s Storm-3358 reporting is explicitly about Southeast Asian finance [6]. And CL-STA-1238’s target list covers the semiconductor supply chain layer by layer, including materials, photomask and standards work [1]. Unit 42 does not name victim countries beyond the US and Canada, so regional exposure is an inference, not a finding.
What to do
- Remove default and shared credentials from internet-facing applications and partner integrations, and enforce strong authentication [1].
- Inventory and close open registration endpoints, unauthenticated cloud-storage write paths and admin interfaces exposed to the internet [1][11].
- Treat any secret found in code or configuration as compromised. Deleting it does not revoke it; rotate it [7][9].
- Prefer managed or short-lived workload identities, set maximum secret lifetimes and apply least privilege to service principals and IAM roles [7][8].
- Hunt for successful calls inside bursts of denied API requests, and for one source touching many service principals in a short window [7][9].
- Alert on machine-speed behaviour: high request variety, rapid retries after errors and fast progression between stages. Microsoft reports that decoys and tarpits reduced agent traffic reaching real assets [6].
- Watch egress for unfamiliar model-proxy or API-reseller traffic, which Censys found co-located with attack consoles [10].
Close
The Cairn case does not show a new class of exploit. It shows a state-aligned operator converting an open-source research tool into a collection pipeline that runs unattended against a supply chain. Defences that already work against password spraying and exposed secrets still work. They now need to hold at machine speed and across every partner that can reach the network [1][12].
Sources
Every R3KONX article cites its primary material. 13 sources, in order of first citation. Links open the original publication.
- Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors Palo Alto Networks Unit 42 · 2026-11-12
- Cairn (project repository and README) GitHub / oritera · 2026-11
- Disrupting the first reported AI-orchestrated cyber espionage campaign Anthropic · 2025-11
- Chinese-Speaking Operator Uses AI Agents to Target Government and Education Systems Across Asia Hunt.io · 2026-09-03
- AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise Microsoft Threat Intelligence · 2026-10-21
- ARTEX abuse expands as cybercriminals target financial sector Microsoft Threat Intelligence · 2026-11-06
- Threat actors use agentic AI tooling to find and abuse exposed secrets Microsoft Threat Intelligence · 2026-11-05
- Hackers use AntiFox AI agents in AWS credential-theft attacks BleepingComputer · 2026-11-11
- AntiFox Targeting AWS Credentials: The Latest Wave of AI Agent Attacks GitGuardian · 2026-11-11
- ARTEX: Agentic Asbestos Censys · 2026-11-05
- Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns Check Point Research · 2026-11-05
- Impact of AI on cyber threat from now to 2027 UK National Cyber Security Centre · 2025-05-07
- AntiFox AI Pentesting Platform Linked to Attacks That Accessed AWS Production Secrets The Hacker News · 2026-11-11
Researched and written by the R3KONX analysis desk from the cited primary material. Attribution statements are the named researchers' own assessments and have not been independently verified; victim organisations are described only as the sources describe them, and exposed-host counts depend on each vendor's scanning method and date. Corrections: event@r3konx.asia.
