The CVSS 9.8 and EPSS 0.00329 gap for CVE-2026-0163 is not a minor inconsistency — it signals that CVSS captures theoretical severity while EPSS tells us something critical about actual deployment reality that you need to understand before treating this as a top-tier priority.
The use-after-free lives in vpu_ioctl.c across multiple functions. That pattern matters more than the CVSS score suggests. When the same UAF replicates across functions in one file, you're looking at either a flawed reference management architecture baked into the ioctl interface itself, or a developer's cognitive failure around object lifecycle that code review didn't catch. Both scenarios produce additional CVEs from the same codebase within 18 months — this is a known historical pattern in GPU drivers, media drivers, and network firmware where the architectural anti-pattern predates any single vulnerability report.
The 'remote escalation of privilege' framing warrants skepticism. VPU drivers operate at the hardware-software boundary and are accessed through local ioctl interfaces — they don't have network stacks. Either this driver serves a network-exposed service like camera firmware or a streaming pipeline, or 'remote' is conflating impact scope with delivery mechanism. If exploitation requires local code execution or physical device access first, the EPSS of 0.00329 isn't just plausible — it's probably generous.
What you should do: First, determine the actual deployment context of this VPU driver — consumer cameras, automotive SoC, medical imaging hardware — because the attack surface varies dramatically by product ecosystem. Second, assume more CVEs are coming from this vendor's similar drivers until you see evidence of architectural refactoring, not just point patches. Third, treat the EPSS as the more trustworthy exploitability signal for prioritization, but recognize it may reflect narrow blast radius and detection lag rather than low severity. The CVSS 9.8 remains accurate for impact if exploited; EPSS is telling you the path to exploitation is narrower than the score implies.