CVE-2026-61625

The CVSS 6.8 and EPSS 0.00303 for CVE-2026-61625 will mislead your prioritization if you treat them at face value. These scores assume an external attacker must probabilistically manipulate S3 or GCS objects to trigger the path traversal in vmrestore. That attacker model is wrong for production deployments—and the three separate release trains being patched (1.122.25, 1.136.12, 1.146.0) confirm this vulnerability was systemic, not incidental. The real threat model treats your backup source as already compromised. If an operator has pivoted to the backup repository—whether through credential theft, insider threat, or supply chain compromise—the path traversal becomes a near-guaranteed write primitive. The CVSS captures file overwrite, but you're not done analyzing. The critical escalation path: if vmrestore runs with elevated filesystem permissions in a containerized environment (common for infrastructure daemons performing backup restoration), this path traversal is a container escape primitive. You're not clobbering config files—you're potentially writing to /etc/cron.d, systemd drop-ins, or kernel module paths if the capability set permits. This is not theoretical; this is the default architecture for monitoring infrastructure at scale. Worse, enterprise backup infrastructure almost universally includes automated restore testing for compliance verification. That scheduled restore job is a remote trigger for the write primitive. You compromise the backup source, and the scheduled restore executes your payload for you. The EPSS treats this as low-probability; the actual exploitability is a scheduled event. Your detection logic likely monitors backup ingestion but not post-restore filesystem anomalies. Your threat model likely trusts backup sources implicitly. Check both. Verify whether vmrestore runs with host filesystem access in your Kubernetes environment, and treat that permission set as your actual blast radius. The vulnerability existed across three parallel release trains because the implicit trust model—'input from this channel is safe because it arrived via this mechanism'—was architectural, not incidental. Other restore pathways in VictoriaMetrics likely carry the same assumption.

Reviewed through automated stages and approved by a human before publication.