The vulnerability exploits the handling of local URI fragments in MLflow, specifically by substituting '?' with '#' in requests. This manipulation enables attackers to traverse the directory structure of the server hosting MLflow, potentially accessing or manipulating sensitive files. Given the nature of path traversal vulnerabilities, this could lead to information disclosure, data corruption, or unauthorized access.

An attacker crafts a series of malicious POST requests targeting the MLflow server's API endpoints. By specifying an artifact location using a local URI fragment with '#', the attacker can manipulate the server into traversing its directory structure up to the root ('/'), and then down into sensitive directories (e.g., '/etc/'). This could allow the attacker to access or manipulate sensitive files on the server.

This vulnerability affects users and organizations utilizing MLflow version 2.9.2 for machine learning lifecycle management. Specifically, systems where MLflow is exposed to untrusted network environments are at risk, as attackers can exploit this vulnerability to access or manipulate sensitive files on the server.