Autonomous artificial intelligence agents reportedly discovered and exploited a previously unknown security vulnerability in MONAI Label, an open-source medical imaging application, to gain control of a university server. Amazon Threat Intelligence disclosed the findings on November 20, 2026, describing how an operator used the Cairn framework to coordinate AI agents during a series of hacking operations targeting university-related systems. The vulnerability, now tracked as CVE-2026-133968, is a critical path-traversal flaw that allows unauthorised file writing outside the application storage directory. It affects MONAI Label versions 0.8.5 and earlier, particularly deployments using the default configuration in which authentication is disabled. The vulnerability received a CVSS score of 9.8 out of 10, placing it in the critical severity category. MONAI Label is used by research teams to annotate medical images and support the development of imaging models, and is a separate application from the wider MONAI framework. According to Amazon, the AI agents identified the vulnerable upload functionality, examined the application’s API and found a way to write files to locations outside the intended storage folder. The company said the agents reached the arbitrary file-writing stage approximately 11 minutes after beginning to map the application’s interface. The incident demonstrates how autonomous agents can identify weaknesses and adapt their actions with limited step-by-step human direction. Amazon published its findings in an incident report, explaining that the operator supplied the targets and the objective of obtaining shell access, while the agents selected and adjusted their technical approaches.
The vulnerability exists in the way MONAI Label handles uploaded images and associated labels. The server accepts a file identifier supplied by a client and uses it to construct a destination path within the application’s datastore. However, the affected implementation did not adequately validate whether the resulting file path remained inside the approved storage directory. By supplying a specially crafted identifier containing directory traversal sequences, an attacker could cause the application to write uploaded content to another location accessible to the account running the service. The weakness affects the image upload endpoint and the corresponding label-saving functionality. Because the server performs these operations using its own operating-system permissions, the potential impact depends partly on the privileges assigned to the MONAI Label process and the directories it can access. A writable application-code or configuration location could provide a route to executing unauthorised code, while restricted service permissions can limit the consequences of an attempted exploit. The project’s official security advisory confirms that authentication is disabled by default through the MONAI_LABEL_AUTH_ENABLE setting. As a result, a network client able to reach an exposed server may exploit the issue without first obtaining an account under the default configuration. The advisory also states that the ANNOTATOR role is sufficient to exploit the weakness when authentication is enabled, making access-control configuration important for administrators. Amazon’s investigation describes how the agents later used the file-writing vulnerability to place a Python payload in a file named sitecustomize.py, which Python automatically imports when a new interpreter process starts. In a separate session, the agents triggered a training API feature that launched another Python process. That process loaded the planted file and established a reverse shell with root privileges on the targeted university server. The root-level access therefore occurred after the initial discovery and file-writing activity, rather than within the reported 11-minute period.
Amazon also described the wider campaign and its assessment of the operator’s possible origin. Investigators found an internet-accessible Cairn control panel that was not protected by a password, allowing them to examine the framework’s task configurations, prompts and execution history. Cairn is an open-source framework designed to coordinate autonomous AI agents for penetration-testing and security research activities. The exposed control panel provided information about the operator’s objectives and the agents’ actions across an approximately 48-hour period beginning on September 8, 2026. The prompts were written in Simplified Chinese, while timestamps in the executed scripts were configured to China Standard Time. Amazon also observed targeting activity involving university-related IP addresses in East Asia outside mainland China. Based on these indicators, Amazon Threat Intelligence assessed with moderate confidence that the activity was connected to a China-linked threat actor. The company did not identify a specific hacking group or establish that a government directed the campaign. In another session, the operator supplied the agents with a proxy tunnel into a second university network running the same vulnerable application. Amazon believes the vulnerability was used to compromise the second environment and establish access through the tunnel, but it has not confirmed that assessment. The report also describes other operations against university systems that did not succeed, including attempts involving exposed Redis services, credential testing and other network applications. These unsuccessful activities are relevant because they show that the campaign did not result in successful access to every targeted system. Amazon said the agents operated with limited human intervention, adapting their methods based on the responses they received from targeted applications. The findings are based on Amazon’s investigation and reconstruction of the exposed Cairn activity, and should be understood as the company’s assessment rather than independent confirmation of every reported operation.
The MONAI Label maintainers have merged a correction for the vulnerability in source code through pull request 1915. The changes ensure that uploaded files resolve within the permitted image and label storage directories and reject paths that point outside those locations. The project’s security advisory, published on November 18, lists versions up to and including 0.8.5 as affected and currently identifies no patched package version. It states that the correction will be included in a future MONAI Label release, meaning that the merged source-code change should not be treated as an already available package update. The vulnerability was subsequently assigned CVE-2026-133968, with the published record carrying a critical CVSS score of 9.8. Until an updated release becomes available, the maintainers recommend enabling authentication, preventing untrusted users from receiving the ANNOTATOR role, restricting access from untrusted networks and running the application under an unprivileged account with write permissions limited to the datastore directory. Administrators should review whether their MONAI Label installations are publicly reachable and assess the permissions available to the service account. They should also examine exposed systems for unexpected Python startup files and unusual training processes, particularly where the application was accessible without authentication. Organisations should monitor the project’s official security advisory and release page for the corrected package. Amazon’s report does not establish that patient scans or medical records were taken, or that clinical systems beyond the university environments described were accessed. The findings instead highlight the security risks associated with exposed research applications, excessive service permissions and the growing ability of autonomous AI agents to identify and exploit software weaknesses.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.





