China-Linked Attackers Use Autonomous AI Agents to Exploit 0-Day in MONAI Label
Suspected Chinese state-sponsored actors leveraged autonomous AI agents powered by the Cairn framework to discover and exploit a zero-day vulnerability in the MONAI Label application, gaining root access to a university server.

Suspected Chinese state-sponsored actors have deployed autonomous AI agents, powered by the open-source Cairn penetration-testing framework, to autonomously discover and exploit a zero-day vulnerability, ultimately gaining root access to a university server. This sophisticated operation, observed by Amazon's Threat Intelligence Team, highlights a significant advancement in AI-driven cyber warfare, where agents operate with minimal to no human intervention.
The campaign, which targeted universities in East Asia outside of Mainland China, unfolded over a 48-hour period beginning September 8. The AI agents initially attempted to breach other university systems through various methods, including exploiting a Redis server, attempting Log4Shell injections, and conducting server-side request forgery (SSRF) attacks. These initial attempts, which involved hundreds of sequential actions and adaptive strategies, failed to yield success, demonstrating the agents' persistence and ability to pivot without human direction.
Following the unsuccessful attempts, the AI agents identified and targeted a zero-day path traversal vulnerability within the MONAI Label application, an open-source tool used for medical imaging AI. The agents autonomously mapped the application's API endpoints, identified a vulnerable upload endpoint, and successfully exploited the path traversal flaw to arbitrarily write files on the server within approximately 11 minutes. This exploit was then leveraged to execute a Python reverse shell with root privileges, granting the attackers full control over the compromised server.
This incident is part of a broader trend of AI agents being adopted for both espionage and criminal operations. Amazon's findings align with previous reports from Gambit Security, which documented a Chinese operator using Cairn to compromise at least 27 companies. In those instances, the attacker also employed other open-source AI tools to exfiltrate sensitive data from various high-profile organizations, including a Fortune 500 hospitality company and a major US airline.
The vulnerability in MONAI Label was reported to its maintainers by Amazon, who subsequently released a fix and a security advisory. This incident underscores the growing threat posed by autonomous AI agents in the cybersecurity landscape. Their ability to independently identify vulnerabilities, craft exploits, and execute complex attack chains presents a formidable challenge for defenders.
The use of Cairn, which has gained traction among cybercriminals, allows for multi-stage intrusions to be carried out with unprecedented autonomy. The prompts used in this specific campaign were in Simplified Chinese, and the timestamps in the executed scripts were configured to China Standard Time, further indicating the origin of the operation.
This event serves as a stark reminder of the evolving threat landscape, where AI is increasingly weaponized. The autonomous nature of these attacks means that response times are critical, and proactive defense strategies must adapt to counter AI-driven threats that can operate at machine speed and scale.
This new report from Amazon's Threat Intelligence team provides further details on the campaign, including the specific AI penetration testing tool, Cairn, used by the attackers. It also highlights that the discovered vulnerability was a path traversal flaw in an open-source medical imaging AI application, which researchers assess with high confidence to be a zero-day due to its unpatched status at the time of discovery.
The new article provides further technical details on the attack, including how the AI agents identified the application's framework, mapped API endpoints, and exploited a path traversal flaw in an upload endpoint. It also elaborates on the method used to achieve persistence: writing a reverse shell payload to sitecustomize.py and triggering a training API endpoint to establish a root-level reverse shell. Additionally, the article notes that Monai has since issued a patch for the zero-day and is requesting a CVE identifier.
The vulnerability, now cataloged as CVE-2026-133968 and rated a critical 9.8, allows unauthenticated attackers to write files outside the designated storage directory due to a path-traversal flaw. While maintainers have merged a fix in the source code, a patched version for installation is not yet available, leaving users to implement manual mitigations such as restricting network access and enabling authentication.
The MONAI Label developers have released an advisory confirming the vulnerability, rating it critical and affecting versions 0.8.5 and earlier. They note that authentication is off by default, making any server accessible over the network exploitable. A fix has been developed and will be included in an upcoming release, providing a clear remediation path for affected users.