Cairn Framework: AI-driven attacks exploiting known vulnerabilities
An operator specifies a target organisation and an overarching objective; the software takes care of the rest. This is how Palo Alto Networks Unit 42 describes a campaign which, between August and September 2026, resulted in unauthorised access or data exfiltration at 10 out of 19 targeted organisations (dig.watch). The tool is called Cairn and is an open-source framework for AI-assisted penetration testing.
Cairn enables AI agents to plan attacks collaboratively
Cairn works with AI agents that store their findings in a shared knowledge base and use this to plan their next steps. The framework maps the attack surface, searches for vulnerabilities, attempts exploits, collects credentials and aims to escalate privileges and gain access to the internal network. It is model-agnostic. In this campaign, the Chinese language models Zhipu GLM 5.2 and DeepSeek V4 drove the agents (Dark Reading).
The campaign
Period: August to September 2026, published by Unit 42 on 12 November 2026
25 operations against 19 organisations, with 10 confirmed instances of unauthorised access or data exfiltration
Sectors affected: Semiconductor supply chain, defence policy, civil society
Attribution: China-linked intelligence gathering, with moderate confidence according to Unit 42
The second documented use of Cairn by China-linked actors
At one organisation, more than one million personal data records were exposed; elsewhere, unauthorised administrator accounts were created in production environments. The attribution is based on circumstantial evidence such as Chinese-language instructions and the selection of targets. It is not a confirmed attribution.
The vulnerabilities exploited were all known
Unit 42 found weak and default credentials, authentication bypasses, exposed cloud APIs, insecure access controls and known vulnerabilities. AI thus increases the volume and scope of the attacks, not their complexity. What these entry points have in common is that they are visible from the outside and can be identified without specialist knowledge. This is precisely what an agent operating as part of a large-scale operation tests: a single operator or a small team was thus able to target 19 organisations simultaneously, whereas previously the number of targets was limited by the number of working hours available.
Questions every organisation should be asking itself right now
Which systems are accessible from the internet? Is there a complete list, and who maintains it?
Are there still default or weak passwords in use? Who checks this regularly?
Are cloud storage and APIs consistently authenticated? Storage accessible without logging in is a preferred point of entry.
Which known vulnerabilities are present on exposed systems? Systems exposed to the internet should be prioritised for patching.
Would anyone notice at night if a new administrator account were created or if an IP address were sending tens of thousands of requests?
Anyone who cannot answer any of these questions with certainty should start there first. Tools such as Cairn search for precisely these vulnerabilities.
Cairn leaves traces that a SOC can detect
In the Unit 42 investigation, Cairn sent hundreds of thousands of high-frequency requests from a single IP address, which can be detected in the central logging system at the network edge (Bright Defense). Further alerts can be derived from the campaign’s results:
Numerous failed login attempts, followed by a successful login to the same application
New administrator accounts outside of planned changes
Access to internal databases or directories by accounts that would not normally use them
Write access to files in cloud storage without authentication
Such rules are only useful if someone is monitoring the alerts round the clock.
If the attack succeeds: the first steps
Restrict access: Reset compromised login credentials immediately.
Check all accounts and tokens created during the breach. Simply changing the original password leaves access points already created by attackers open.
Back up logs before making any changes to systems, so that forensic analysts can reconstruct the sequence of events.
Identify the affected data and check whether reporting obligations under the GDPR or NIS2 apply. The deadlines begin from the moment awareness is gained.
Conclusion: AI makes attacks faster, not smarter
Cairn demonstrates what a small team can achieve with autonomous software: targeting many systems simultaneously and succeeding where there were already vulnerabilities. Those who have a firm grasp of identities, attack surfaces and patch statuses can prevent such campaigns from gaining a foothold.
For those wishing to check whether detection and escalation systems would recognise such an attack in a real-world scenario, we recommend our Incident Drill. It simulates real-world attack paths based on the MITRE ATT&CK framework, including credential access and privilege escalation, in a controlled environment.
FAQ
What is the Cairn framework?
An open-source framework for AI-assisted penetration testing. AI agents handle reconnaissance, vulnerability scanning and exploitation. The attackers used a customised version.
Who is behind the attacks?
Unit 42 is tracking the campaign as CL-STA-1238 and, with moderate confidence, attributes it to Chinese intelligence gathering. It is likely that a single operator or a small team was responsible.
Which organisations were affected?
19 organisations were attacked; in 10 cases, there was confirmed access or data exfiltration. They are drawn from the semiconductor supply chain, the defence sector and civil society. Unit 42 does not name them.
Do AI-driven attacks differ technically from traditional ones?
According to Unit 42, hardly at all. The attackers exploited known vulnerabilities; the AI increases the speed, scale and parallelism of the attacks.
What measures help?
Identity hygiene, a smaller attack surface, consistent patching, centralised logging with alerts and a tested incident response plan.
