The CVSS-EPSS divergence in CVE-2026-65582 should concern you more, not less. A 7.7 score paired with an EPSS of 0.00424 creates the appearance of manageable risk, but this metric gap is characteristic of a vulnerability class that threat models systematically underweight: authenticated arbitrary file download in multi-tenant contexts.
The subscriber role in AI Hub 1.3.10 is the analytical tell. Multi-tenant file access endpoints require developers to reason about cross-tenant isolation at every operation—a cognitively demanding task that most frameworks don't primitive cleanly. The result is predictable: code that validates "is this a subscriber?" but skips "is this subscriber allowed to access this specific file?" This isn't developer negligence; it's a structural gap between authentication as a gate and authorization as a resource-level check. The fix often requires retrofitting authorization checks into code paths that were never architected to contain them.
The low EPSS reflects what this vulnerability is not—a flashy initial-access flaw that mass scanners catch. It reflects what it actually is: a post-compromise pathway. Attackers who compromise an adjacent system (a CI pipeline, a data science workstation, a third-party integration) can pivot through any valid subscriber context to access what AI Hubs actually store—training datasets, model weights, proprietary pipelines, and API credentials. This is exploitation as force multiplier, not as entry point. EPSS modeling doesn't capture it well because the class operates below its detection threshold.
Compounding the risk: these endpoints are often forgotten code paths, written during initial architecture under optimistic assumptions about subscriber count and data sensitivity, then left static as the product evolved and the data they guard became mission-critical. The security assumptions embedded in the original implementation now govern access to production ML infrastructure.
Check whether AI Hub 1.3.10 is within a supported patch window. If it's at end-of-life or the vendor timeline exceeds your remediation SLAs, treat this as a compensating-controls priority—network segmentation, subscriber-access logging, and heightened monitoring for lateral movement patterns through the file download endpoint. The compound risk is this vulnerability sitting dormant in environments already partially compromised, awaiting activation.