The era of human-in-the-loop cyber warfare is officially over. Amazon CISO CJ Moses disclosed that suspected Chinese state-sponsored actors deployed autonomous AI agents to locate and exploit a zero-day vulnerability in an open-source medical imaging application, successfully compromising a university server. This incident marks a significant escalation in the use of AI-driven penetration testing frameworks for espionage, moving beyond scripted attacks to dynamic, adaptive intrusion campaigns that operate with minimal human oversight.

The Cairn Framework and Autonomous Adaptation

The attackers utilized Cairn, an open-source penetration-testing framework that has gained traction among cybercriminals for executing multi-stage intrusions without human assistance. According to Moses, the agents operated autonomously for nearly a full day, adapting their strategies in real-time. The campaign began on September 8 and lasted 48 hours, with prompts written in Simplified Chinese and scripts timestamped in China Standard Time. Amazon’s Threat Intelligence Team discovered the operation on September 10 by identifying a publicly hosted instance of Cairn that provided visibility into the targeting of East Asian universities outside Mainland China.

From Failed Attacks to Zero-Day Discovery

Before uncovering the critical zero-day, the AI agents engaged in a grueling 16-hour sequence of 148 failed attempts against a different university’s Redis server. These attempts included abusing Redis replication features, escaping scripting sandboxes, and attempting Log4Shell injections and SSRF attacks against co-located services. A separate operation involving a GPU server saw the agents compiling a list of 80 real user names from public academic papers for credential sprays. Despite these initial failures, the agents’ ability to pivot and adapt without direct human direction highlighted their evolving sophistication.

Exploiting MONAI Label via Path Traversal

The breakthrough occurred when the agents targeted MONAI Label, an open-source medical imaging AI application. The AI framework identified the application’s structure, mapped its API endpoints, and detected an upload endpoint with a vulnerable filename parameter. Within approximately 11 minutes of mapping the API, the agents exploited a path traversal vulnerability to write arbitrary files on the server. Subsequently, they leveraged the same flaw to execute a Python reverse shell with root privileges, gaining full control of the system. Amazon notified the project maintainers, who quickly released a security advisory and a fix.

Key Takeaways

  • Autonomous AI agents can now perform complex, multi-stage attacks that previously required skilled human operators.
  • The Cairn framework is being adopted by both espionage and criminal groups for its ability to operate without constant human input.
  • Traditional defenses relying on static signatures or predictable attack patterns are increasingly ineffective against adaptive AI agents.

The Bottom Line

This isn't just a security update; it's a paradigm shift. When AI agents can autonomously map APIs, identify zero-days, and pivot from failed attacks to successful root access in under a day, the asymmetry of cyber warfare changes drastically. We are no longer just fighting humans with tools; we are fighting tools that think like humans, but never sleep.