Hackers use autonomous AI tool Cairn to find zero-day, take over server
A Chinese-linked threat actor used the autonomous AI tool Cairn to exploit a MONAI Label zero-day vulnerability and seize root access on an imaging server.
By Dillip Chowdary • Nov 20, 2026 • Source: BleepingComputer

At a glance
- Attack tool
- Cairn autonomous AI agent framework
- Vulnerable software
- MONAI Label medical imaging application
- Patch date
- November 18, 2026
- Attack dates
- September 8–10, 2026
- Campaign targets
- East Asian educational institutions outside China
- Attributed actor
- Chinese-nexus threat group
In this post
A suspected Chinese-nexus threat actor deployed an open-source autonomous AI framework named Cairn to discover a zero-day flaw in a medical imaging application, as documented in BleepingComputer's report. After operators instructed the agents to obtain shell access, the software autonomously mapped the application and seized full root control of an enterprise server. Amazon Threat Intelligence uncovered the campaign because the attackers left their management instance publicly accessible without authentication.
This report details the mechanics of autonomous multi-agent cyberattacks, the operational boundaries observed during live intrusions, and the defensive measures required to protect infrastructure. It is designed for system administrators, security engineers, and defenders who evaluate threat telemetry, indicators of compromise, and automated response capabilities.
01Exposed controller and initial prompt instructions
The attacks occurred between September 8 and 10, 2026, primarily targeting educational institutions across East Asia outside China. Amazon Threat Intelligence discovered the intrusion after finding that the threat actors left their Cairn platform publicly accessible without authentication. That operational oversight exposed the entire campaign, granting researchers direct visibility into the active console logs and command history.
The threat actors directed the autonomous agents with a simple high-level objective to obtain shell access, avoiding manual exploit scripting. The agents then operated independently to discover network paths, analyze target surfaces, and execute tactical decisions. Console logs confirmed that the multi-agent framework handled task distribution, target exploration, and exploitation attempts without requiring continuous human guidance.
Amazon attributed the campaign to a Chinese-nexus threat actor with moderate analytical confidence. Investigators based this finding on the language used in the retrieved operator prompts, the activity timestamps, and the targeting of regional universities. The exposed infrastructure provided clear operational proof of how hostile groups now deploy autonomous agents against public institutions.
02Multi-agent penetration software and prior campaigns
Cairn is an open-source penetration-testing platform engineered to coordinate multiple AI agents during complex offensive cyber-operations. Created legitimately to assist security researchers and strengthen defensive postures, the software manages distributed agents that autonomously evaluate network targets. However, threat actors have increasingly repurposed the framework to conduct automated, unauthorized attacks against external organizations.
Hostile deployments of Cairn have expanded across multiple sectors in recent months. In September 2026, Gambit Security documented threat actors abusing the framework during a large-scale cybercrime operation designed to steal credit card data. That incident demonstrated that financial adversaries could effectively integrate autonomous AI tooling to streamline high-volume credential and data harvesting.
Google subsequently reported that threat actors deployed Cairn in a targeted cyberattack against an aerospace organization in the United States. That deployment highlighted the appeal of autonomous frameworks for sophisticated actors targeting critical industrial and defense infrastructure. The platform enables operators to automate repetitive reconnaissance and tactical decision-making across complex victim environments.
03Sixteen-hour credential attacks and operational limits
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
Operational telemetry gathered during the campaign revealed that the autonomous AI agents were not uniformly successful across all target systems. During one engagement targeting a Redis server, the agents sustained an intrusion effort lasting 16 hours. Despite executing 148 distinct autonomous actions during that window, the software failed to breach the host or establish administrative access.
In another separate intrusion directed at a university graphics processing unit server, the agents attempted an exhaustive brute-force attack. Over a six-hour period, the autonomous software generated and tested more than 30,000 credential combinations against the GPU cluster. The server resisted the sustained volume of queries, leaving the agents unable to secure an initial foothold.
These failed operations demonstrate the practical constraints of current autonomous agent frameworks when attacking hardened infrastructure. When confronted with systems lacking easily exploitable flaws or exposed APIs, the agents relied on brute-force repetition rather than creative evasion. These boundaries illustrate that autonomous tooling still depends heavily on existing software vulnerabilities to achieve meaningful penetration.
04Eleven-minute medical imaging vulnerability exploit chain
The agents achieved their primary breakthrough by discovering a zero-day vulnerability in MONAI Label, an open-source medical imaging application that uses AI to help researchers annotate images. The software uncovered a previously unknown path traversal flaw that permitted arbitrary file writes. Within eleven minutes, the autonomous agents progressed from mapping the application API to full exploitation.
In a subsequent session, the agents used the path traversal flaw to plant a Python reverse shell inside sitecustomize.py, a file that executes automatically whenever Python starts. The agents then triggered an application API endpoint, running the payload to establish root access on the server. The entire sequence executed automatically without requiring real-time exploit modification from human controllers.
Following the initial compromise, the human operator supplied the agents with proxy access into a second university network running the same MONAI Label software. The controller then instructed the autonomous agents to target additional machines across the newly accessible environment. This handoff demonstrated how human operators can leverage successful automated zero-day discoveries to pivot rapidly across distinct victim networks.
05Vendor security update and defensive automation
Amazon Threat Intelligence promptly notified MONAI Label maintainers regarding the active path traversal vulnerability discovered during the campaign. The development team validated the defect and released an official security update to address the flaw on November 18, 2026. Administrators running MONAI Label instances must apply this patch immediately to eliminate the arbitrary file write risk.
Amazon also published concrete threat indicators observed between September 8 and 10, 2026, to assist defensive teams with threat hunting. Investigators identified two external IP addresses tied to the campaign: 142[.]4.98.145 and 142[.]4.107.34. Security analysts should search gateway and firewall logs for historical traffic involving these network addresses to identify potential reconnaissance or compromise.
Amazon warned that autonomous AI agents can identify and exploit zero-day vulnerabilities far faster than human defenders can analyze logs and respond. The rapid eleven-minute exploit progression underscores the growing challenge facing traditional security operations centers. Consequently, organizations must deploy automated defensive capabilities and continuous configuration auditing rather than relying solely on manual investigation workflows.
Frequently asked questions
What is Cairn and how was it used in this attack?
Cairn is an open-source penetration-testing platform designed to coordinate multiple autonomous AI agents. Attackers gave the agents a high-level instruction to obtain shell access, allowing the software to discover and exploit a zero-day vulnerability independently.
Which application contained the zero-day vulnerability?
The vulnerability affected MONAI Label, an open-source medical imaging application that uses artificial intelligence to assist researchers with image annotation. The path traversal flaw enabled arbitrary file writes that attackers leveraged to gain root server access.
Has the MONAI Label vulnerability been patched?
Yes, MONAI Label maintainers released a security update addressing the path traversal flaw on November 18, 2026, after receiving a disclosure report from Amazon Threat Intelligence.
What operational failures did the AI agents experience during the campaign?
The agents failed during a 16-hour, 148-action attempt to compromise a Redis server. Additionally, they tested more than 30,000 credential combinations over six hours against a university GPU server without gaining access.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement