Linux Kernel CVE-2026-74419

The target system must be running a Linux kernel version that includes the vulnerable amdxdna driver (commit 850d71f6bf4c present, fix commits absent) on hardware equipped with an AMD XDNA NPU (Ryzen AI-series processor). The attacking user must hold at least low-privilege local access (interactive shell or equivalent) and the operating system must grant access to the /dev/accel device node for the user's account - typically via group membership configured by udev rules. No special kernel configuration flag is required beyond the driver being loaded. Exploitation is limited by the hardware requirement (AMD XDNA silicon only), the need for a local shell, and device-node permissions that are typically restricted on hardened or server deployments.

Real-world risk is moderate-to-low despite the 7.3 CVSS score. The EPSS score of 0.14% (4th percentile) indicates a statistically very low likelihood of exploitation in the near term, which is consistent with the driver's niche hardware dependency - the amdxdna driver only loads on systems with AMD XDNA/NPU silicon (Ryzen AI). The CVSS vector (AV:L/AC:L/PR:L) correctly reflects that exploitation requires local shell access on an affected machine, and the device node must be accessible to the calling user, typically through group membership or udev rules. KEV status is negative - no active exploitation confirmed. There is no public POC. The C:H/I:H rating is technically coherent: a copy_to_user overflow from kernel can spill adjacent kernel data into user space (confidentiality) and overwrite adjacent user-space memory (integrity), making this more dangerous than a pure info-leak. Prioritization should weight hardware prevalence: organizations with AMD Ryzen AI hardware in their fleet (developer laptops, AI workstations) should patch promptly; environments without this hardware class are not affected.

A local user with low privileges on a workstation equipped with an AMD Ryzen AI processor opens the amdxdna accelerator device node and issues an ioctl with a carefully sized destination buffer smaller than the kernel's internal data structure. The kernel copies beyond the application buffer, spilling adjacent kernel data into user space or overwriting adjacent user-space memory, which an attacker could leverage to leak kernel addresses or corrupt process memory. No public proof-of-concept exploit exists at time of analysis.