Unit 42 uncovers Cairn AI breaches at 10 organizations after researchers traced 25 autonomous intrusion operations against at least 19 targets. The August and September campaign concentrated on semiconductor manufacturing, defense policy and civil-society groups.

 

Palo Alto Networks’ Unit 42 said Chinese-speaking attackers paired the open-source Cairn penetration-testing framework with Zhipu GLM and DeepSeek models. The researchers assessed with moderate confidence that the activity was China-linked intelligence collection, rather than making a definitive attribution.

 

The investigation establishes four important findings:

  • Cairn received goals, planned attacks and recorded results autonomously.
  • Ten organizations suffered unauthorized access or data retrieval.
  • Known identity and configuration flaws delivered the confirmed results.
  • Operations lasted from 20 minutes to more than 50 hours.

 

Further Reading

 

Unit 42 Uncovers Cairn AI Breaches at 10 Organizations

The Unit 42 investigation, published November 12, describes 25 operations conducted by what researchers believe was one operator or a small team. Seven targets were in the semiconductor supply chain, three involved defense or defense policy, and four were civil-society organizations.

 

Researchers confirmed unauthorized access or data retrieval at 10 targets. At one industry standards body, attackers obtained dozens of valid credentials and reached non-public directory data. They also changed a live standards-ballot document through an unauthenticated cloud-storage write.

 

At a semiconductor company, the framework brute-forced a production credential that worked across multiple accounts. It retrieved application and user data, then created unauthorized administrator accounts and a persistent API token, giving the operators a durable path back into the production system.

 

A teachers’ union suffered exposure involving more than one million member records, about 31,000 customer-management contacts and staff password hashes. A separate defense-policy program exposed roughly 500 user records, including accounts tied to U.S. military and government domains.

 

How Cairn Turned One-Line Goals Into Intrusion Plans

Cairn is a public AI penetration-testing framework designed to coordinate multiple tasks around a shared record of facts and intended actions. In the observed campaign, operators supplied a target domain and a short objective; the framework mapped the attack surface and divided the goal into smaller steps.

 

The same AI worker handled reconnaissance, web applications, services, credentials, lateral movement and persistence. It repeatedly reviewed confirmed findings, selected another action and stored the outcome, stopping only when it judged the objective complete, exhausted its options or was halted by the operator.

 

This matters because the framework could run different operations simultaneously. Unit 42 said the campaign pursued 19 organizations within two weeks, reducing the operator time that would normally constrain targeted intrusion work.

 

The automation did not eliminate human preparation. In some cases, attackers supplied target-specific information, including VPN gateway details and credentials gathered beforehand. Cairn accelerated exploration and exploitation after receiving that starting intelligence.

 

GLM and DeepSeek Models Powered 25 Operations

Unit 42 counted 13 operations using Zhipu GLM 5.2, eight using DeepSeek V4 Flash and four using DeepSeek V4 Pro. The researchers believe the attackers compared models and used the higher-tier DeepSeek model for repeated attempts against targets that resisted the Flash version.

 

The attackers also loaded a custom skill package that added scanning and exploitation functions beyond the public Cairn project. Their prompts portrayed every task as an authorized security assessment and instructed the worker not to refuse, while generated authorization files repeated the same false premise.

 

Unit 42 said neither model rejected that framing. The finding shows a practical guardrail problem: an agent can receive an authorization claim from the same untrusted operator requesting the action, then treat the claim as evidence that the work is legitimate.

 

The report does not show that the models invented a new class of exploit. Instead, they increased the number of systems an attacker could examine and the time each operation could continue without constant human direction.

 

Known Identity Flaws Delivered the Campaign’s Results

The confirmed paths included weak or default passwords, authentication bypasses, exposed cloud interfaces, insecure object references, mass-assignment privilege escalation and older public vulnerabilities. These are familiar security failures, but autonomous orchestration tested them across targets at sustained scale.

 

That distinction shapes the defensive response. Organizations can reduce exposure by enforcing phishing-resistant multifactor authentication, removing unused public endpoints, fixing authorization checks, restricting administrator creation, rotating shared credentials and alerting on bulk access or persistent tokens.

 

Unit 42’s moderate-confidence attribution rests on Chinese-language tasking, exclusive use of Chinese models and a target set aligned with established Chinese intelligence priorities. Those signals support the assessment, but they do not publicly identify a government sponsor or named threat group.

 

The disclosure arrives as U.S. officials discuss stronger reporting for AI-enabled cyber incidents and vulnerabilities. Reuters reported that a federal AI task force considered information-sharing standards for developers and critical-infrastructure operators on November 12.

 

Cairn’s significance is therefore operational rather than theoretical. Commodity model access and a public agent framework let a small team pursue many targets, while ordinary identity weaknesses still determined which attempts became real compromises.

 

Illustrative feature photograph: semiconductor cleanroom wafer handling by Niftyfiftycat, via Wikimedia Commons, licensed under CC BY-SA 4.0. It does not depict any targeted organization or intrusion. The original 16:9 image was resized.