Attackers Exploit MLflow SSRF Flaw to Steal Cloud Credentials and Secrets
AI-summarised brief · reviewed before publication
Researchers at watchTowr and VulnCheck reported active exploitation of two critical flaws disclosed in mid‑August 2026. CVE‑2026‑64849 in the open‑source AI platform MLflow enables attackers to misuse the model‑registry webhook to proxy requests to internal cloud metadata services, allowing them to harvest cloud credentials and secrets. Within hours of the CVE’s assignment, malicious actors began scanning the internet for exposed MLflow instances, and global honeypot data shows attempts to extract secrets from known internal IP ranges. A separate issue, CVE‑2026‑25895, affects the web‑based SCADA/HMI tool FUXA; a single IP has been seen probing roughly 60 publicly exposed installations, trying to overwrite the main.js file via a path‑traversal vector, though no remote‑code execution payloads have been observed yet. Organizations are urged to patch, audit logs, and verify that credentials have not been compromised.
💡 Why It Matters
- · The MLflow flaw gives attackers a direct line into cloud environments, turning a seemingly innocuous webhook into a credential‑stealing conduit, while the FUXA scans highlight the growing focus on OT software as a low‑hanging target for supply‑chain attacks.
