Threat actors utilized the open-source AI penetration-testing framework named "Cairn" to locate a previously unknown flaw in a medical imaging application and subsequently achieved complete server control.
The investigation by Amazon Threat Intelligence commenced on September 10, 2026, after researchers detected an instance of Cairn that was accessible publicly without requiring authentication. This exposure allowed observers to view the platform's operational history, including the prompts issued by the operator and the resulting AI output. The campaign, which targeted educational institutions outside of China in East Asia, was conducted by a Chinese-nexus threat actor between September 8 and September 10, 2026.
Cairn is designed as a legitimate tool for strengthening defenses and research, capable of coordinating numerous AI agents in cyber operations. However, its simplicity of use makes it appealing to adversaries. In this specific engagement, the threat actor instructed the AI agents merely to "obtain shell access," allowing the agents to independently devise and execute methods to uncover and leverage the security defect. According to the researchers, the AI agents autonomously planned, carried out, and modified their strategies as they gained experience, without the human operator dictating every subsequent action.
The operational tempo demonstrated by these AI agents is a key concern; an agent successfully finding and exploiting a zero-day in minutes can surpass the manual triage capacity of many security operations centers. During the campaign, the agents faced several setbacks. For instance, an endeavor spanning 16 hours and involving 148 distinct actions to gain control of a Redis server ultimately failed. Furthermore, in a separate instance targeting a university GPU server, the agents tested over 30,000 credential combinations across six hours but were unsuccessful in brute-forcing entry.
The successful exploitation occurred when the agents identified a path traversal zero-day within MONAI Label, an open-source medical imaging application that employs AI for image labeling assistance. This vulnerability permitted the agents to write files to any location on the server. Within eleven minutes, the agents progressed from mapping the application's API to successfully exploiting the weakness. In a subsequent phase, the AI agents utilized this flaw to write a Python reverse shell into the sitecustomize.py file, which loads automatically when Python initializes, and then triggered an API endpoint to attain root privileges on the server.
Following the initial discovery, the operator provided the agents with a proxy connection into a second university network that utilized the identical vulnerable software, directing them to target additional machines. Amazon Threat Intelligence subsequently informed the maintainers of MONAI Label, leading to the release of a security patch on November 18 to remedy the issue. The researchers assessed the campaign's attribution to a Chinese-nexus threat actor with moderate certainty, basing this on the language present in the retrieved prompts, the timing of the activity, and the selection of targets.
Amazon also disclosed two IP addresses linked to the campaign, observed during the September 8 to September 10, 2026 window. The warning issued by Amazon underscores that AI agents possess the capacity to locate and exploit vulnerabilities more rapidly than security teams can respond, thereby advocating for increased reliance on automated defensive measures over manual investigative processes.
This incident highlights the rapid maturation of offensive capabilities, where generalized, agentic tools like Cairn allow threat actors to automate complex attack chains with minimal human intervention. This development aligns with broader industry observations regarding the increasing sophistication of AI-backed cyber threats, as other entities have noted the deployment of similar autonomous frameworks in various attacks.
Why it matters
This incident highlights the escalating threat of autonomous AI in offensive cyber operations, capable of discovering and exploiting vulnerabilities faster than manual security teams. Operators must recognize that AI-driven attacks shift the focus from simple exploitation to managing rapidly evolving, self-directing threats.






