MLflow & FUXA Flaws Lead to Cloud Credential Theft
Introduction
The security community has uncovered two high‑severity weaknesses that affect widely used open‑source software. One targets MLflow, a popular framework for managing machine‑learning experiments, while the other impacts FUXA, a web‑based interface for supervisory control and data acquisition in industrial settings. Both flaws are already being probed by automated scanners and malicious actors aiming to exfiltrate cloud credentials and proprietary operational data.
Independent analyses from watchTowr and VulnCheck confirm the existence of these issues and outline the tactics used by attackers to turn the vulnerabilities into credential‑theft vectors.
Critical Flaws in MLflow and FUXA
MLflow, widely adopted for tracking experiments, logging metrics, and storing model artifacts, contains a server‑side request forgery (SSRF) flaw. The vulnerability stems from a component that accepts a URL parameter to retrieve files or metadata without proper validation. An attacker can craft a request that forces the MLflow server to contact internal cloud metadata endpoints, such as the instance‑metadata service, and retrieve authentication tokens. Those tokens can then be used to access cloud storage, virtual machines, or other services, effectively turning the platform into a conduit for credential theft. The impact is amplified because many organizations grant broad permissions to their MLflow instances, making the stolen credentials highly valuable.
FUXA, an open‑source web interface designed for supervisory control and data acquisition in industrial environments, suffers from a different class of weakness. Inadequate sanitization of user‑supplied input in its API endpoints enables remote code execution and arbitrary file read operations. By sending specially crafted requests, an adversary can force the application to disclose configuration files that contain usernames, passwords, and API keys used to control PLCs, RTUs, and other field devices. In many OT deployments, these secrets grant direct influence over production processes, turning a software flaw into a potential safety hazard.
Active Exploitation Campaigns Targeting AI and OT Platforms
Threat intelligence feeds indicate that automated scanning tools are already probing the internet for exposed MLflow and FUXA endpoints. Once a vulnerable instance is identified, attackers deploy scripts that exploit the SSRF or RCE vectors, then pivot to cloud credential harvesting or OT configuration extraction. The stolen data is subsequently used to launch lateral movement, ransomware deployment, or sabotage of manufacturing lines. The speed at which these exploits spread underscores the urgency for organizations to remediate the issues promptly and to monitor for anomalous traffic patterns that may indicate ongoing abuse.
Mitigation strategies revolve around tightening network boundaries, applying vendor patches, and enforcing least‑privilege access. For MLflow, administrators should disable outbound HTTP requests to metadata endpoints, employ outbound firewall rules, and rotate any credentials that may have been exposed. For FUXA, upgrading to the latest release, disabling unnecessary API endpoints, and implementing strong authentication mechanisms such as multi‑factor authentication are essential steps. Continuous monitoring with behavior‑based detection can flag suspicious request patterns before they result in data exfiltration.
Grivyonx Expert Analysis
The convergence of AI‑driven development pipelines and critical infrastructure software creates a novel attack surface. While MLflow’s SSRF flaw primarily endangers cloud credentials, its exploitation demonstrates how seemingly benign automation tools can become vectors for large‑scale credential harvesting. Similarly, FUXA’s RCE risk illustrates that OT environments, traditionally isolated, are now exposed to the same internet‑facing threats as traditional IT systems. Both cases reinforce the need for a zero‑trust architecture that treats every service, regardless of its domain, as potentially compromised. Implementing automated patch management, continuous vulnerability scanning, and strict egress filtering can dramatically reduce the attack surface. Moreover, integrating AI‑based anomaly detection into security operations can surface subtle indicators of credential misuse that traditional rule‑based systems might miss, providing an extra layer of protection for enterprises that rely heavily on MLflow and FUXA.
Conclusion
In summary, the recent exploitation of critical flaws in MLflow and FUXA highlights the growing intersection of artificial intelligence, industrial automation, and cyber risk. Organizations must act swiftly to patch the identified vulnerabilities, harden their network perimeters, and adopt proactive monitoring to prevent credential theft and operational disruption. As the threat landscape evolves, partnering with a trusted security and automation platform can help safeguard both cloud‑native AI workloads and industrial control systems, ensuring continuity and resilience in an increasingly connected world.
- #MLflow
- #FUXA
- #SSRF
- #AI Platform
- #OT Security
- #Cloud Credentials
- #Vulnerability Exploitation
