Amazon’s threat intelligence division has identified a campaign in which a China-linked threat actor utilized autonomous artificial intelligence agents to discover and exploit a previously unknown vulnerability in MONAI Label. MONAI Label is an open-source framework frequently used for AI-assisted medical image annotation. The tech giant detected the activity in September after discovering an unsecured instance of Cairn, an open-source AI penetration testing platform.
Because the Cairn instance lacked authentication, researchers were able to observe the operator’s direct prompts and the subsequent output generated by the AI agents. The logs revealed a 48-hour offensive operation directed at universities located in East Asia, specifically outside of mainland China. The operator provided the agents with a list of targets and clear instructions to obtain shell access to the systems.
According to Amazon, the AI agents demonstrated significant speed in their execution. Within approximately 11 minutes of mapping the target application’s API, the agents successfully identified a path traversal vulnerability within a file upload function. This flaw allowed the agents to write arbitrary files to the server. In a subsequent stage of the attack, the agents leveraged this access to plant a Python reverse shell, which was configured to execute automatically whenever the Python interpreter started, ultimately granting the attackers root-level control.
Amazon has confirmed that the vulnerability was a zero-day at the time of the attack, noting the complete absence of any prior security advisories, patches, or mentions in vulnerability databases for the affected software. The MONAI Label development team has since issued an advisory confirming the existence of the critical flaw, which impacts versions 0.8.5 and earlier. The developers noted that because authentication is disabled by default in these versions, any user with network access to the server could potentially exploit the vulnerability.
The investigation also uncovered evidence of lateral movement. After compromising an initial target, the agents were instructed to use a proxy to tunnel into a second university network. Amazon observed that the server reached via this tunnel was running the same vulnerable software, leading investigators to conclude that the threat actor had previously used the same zero-day exploit to compromise that secondary system and establish the network bridge.
Amazon attributed the campaign to a Chinese-nexus threat actor with moderate confidence. This assessment is based on several factors, including the specific language used in the operator’s prompts, the time zone settings observed in the execution scripts, and the strategic nature of the targeted academic institutions.
The MONAI Label project has developed a fix for the vulnerability, which is scheduled for inclusion in the upcoming software release. Organizations utilizing versions 0.8.5 or earlier are advised to review the official project advisory and apply the necessary updates as soon as they become available to mitigate the risk of unauthorized file access and remote code execution.






