· ai-devtools

Amazon reports Chinese hackers used Cairn AI tool to find zero-day in medical imaging app

The discovery of a zero-day vulnerability in a medical imaging application, exploited by autonomous AI agents, reveals a new frontier in offensive cyber capabilities.


Threat actors leveraged the open-source AI penetration-testing framework, Cairn, to uncover and exploit an unknown flaw, thereby gaining complete command over a server.

Amazon Threat Intelligence uncovered this campaign after noticing that the attackers had left their Cairn instance accessible to the public without requiring authentication, which exposed their hacking activities. According to the researchers, this Chinese-nexus threat actor initiated the attacks between September 8 and 10, 2026, primarily aiming at educational institutions across East Asia that were not located in China.

Cairn itself is a legitimate tool designed for defense strengthening and research, enabling the coordination of numerous AI agents in cyber operations. Previously, Gambit Security had documented instances of Cairn misuse during a large-scale operation focused on stealing credit cards in September, and more recently, Google reported its deployment against an aerospace organization within the United States.

During the investigation conducted by Amazon, the AI agents were not consistently successful in their endeavors. For instance, a sustained effort involving 148 actions over a period of 16 hours to seize control of a Redis server ultimately failed. In another instance targeting a university's GPU server, the agents tested over 30,000 credential combinations across six hours but could not successfully brute-force entry.

The successful breach occurred when Cairn's AI agents identified a path traversal zero-day within MONAI Label, an open-source application for medical imaging that utilizes AI to assist researchers in labeling medical images. This vulnerability permitted the agents to write files to any chosen location on the server; within just 11 minutes, the agents progressed from mapping the application's API to successfully exploiting the weakness. Subsequently, in a following session, the AI agents utilized this flaw to deposit a Python reverse shell into the sitecustomize.py file, which automatically loads when Python starts, and then triggered an API endpoint to establish root access to the server.

Following the discovery of the zero-day, the operator furnished the agents with a proxy into a second university network running the identical vulnerable software, subsequently directing them to attack other machines. The researchers reported the flaw to the maintainers of MONAI Label, who subsequently issued a security update to remedy the issue on November 18. The researchers assigned a moderate level of confidence to attributing the campaign to a Chinese-nexus threat actor, basing this assessment on the language used in the retrieved prompts, the timing of the activity, and the selection of the targets. Amazon also disclosed two IP addresses linked to the campaign, which were observed between September 8 and 10, 2026.

Amazon issued a stern warning that the speed at which AI agents can locate and exploit vulnerabilities surpasses the reaction time of security teams, strongly advising organizations to bolster automated defenses rather than relying solely on manual investigations. Adaptive Security's AI agents are capable of analyzing email intent, context, and behavior to uncover invoice fraud, phishing, and business email compromise.

CJ Moses, CISO of Amazon Integrated Security, explained that Cairn is particularly advantageous for attackers because, once given a starting point and a goal articulated in simple language, it can manage multiple AI agents to achieve the objective. In this specific incident, the attacker supplied IP addresses and instructed Cairn to “obtain shell access,” meaning they sought a method to secure sufficient system access for various actions. Moses elaborated, stating, “Based on those inputs, the AI agents planned, executed, and adjusted their approach as they learned, without the operator choosing each next step.”

The campaign involved several operations of varying complexity and outcome, demonstrating the persistence and capability of these autonomous AI agents. The most intensive operation targeted an unauthenticated Redis instance at one university. Over approximately 16 hours and 148 sequential actions, the agents worked through numerous known methods for gaining control of an open Redis server, including abusing its replication feature to load unauthorized code, escaping its scripting sandbox, modifying its configuration files, and corrupting its memory. Additionally, the agents attempted Log4Shell injection against another service running on the same system.

In a separate matter, the FBI released an advisory a week prior concerning the AI tool Artex, noting that it possesses the capacity to transform any large language model into a potent hacking instrument capable of “automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration.” Artex has been linked to multiple intrusions targeting financial institutions in Japan and South Korea, while Microsoft indicated that Chinese state-backed actors utilized Artex in a recent supply chain attack against the IT management software company ConnectWise. Moses added last week that Artex had been observed in credential theft attacks targeting AI models and in operations aimed at cryptocurrency exchanges.

Google researchers also brought Cairn into the spotlight on Tuesday in a report detailing the evolving use of autonomous, multi-agent offensive frameworks, such as CyberStrikeAI, PentAGI, and Artex. Google observed Cairn being employed in intrusions against entities within the U.S. aerospace sector, confirming what AWS found: the tool excels at commencing with a target's IP address and concluding with a defined operational objective. The AWS report on Cairn references a September study from the cybersecurity firm Gambit Security, which documented a financially motivated hacking group using Cairn to compromise online retailers. Eyal Sela, Director of Threat Intelligence at Gambit Security and the report's author, told Recorded Future News that Cairn holds greater significance than Artex because it functions as a general-purpose utility. Sela commented, “While it is good at solving ‘red team’ or capture-the-flag tasks, it can also solve other problems and the main point is that it has no internal structuring for cybersecurity.” He further explained, “When a tool is more generic, it is less predictable because it has less structuring around the way it solves a task.”

Why it matters

This incident demonstrates how autonomous AI tools can rapidly discover and exploit previously unknown vulnerabilities in software. Developers must recognize that AI agents can accelerate the discovery phase of cyberattacks far beyond manual testing capabilities.

Earlier coverage

Related coverage

Also worth reading

  1. next.js 16.5.0-canary.59 released
  2. Bell Canada confirms agentic hacker stole enterprise customer data on November 17 2026
  3. Cyber leaders find clear communication a front-line defense capability
  4. Meta releases Triton TLX, a Native GPU Compiler Extension for advanced hardware control
  5. AI agents autonomously find and exploit 0-day in MONAI Label to take over university server