Linux Kernel EUVD-2026-86975

Requires local access to the AMD XDNA accel device (amdxdna driver), which is only present on systems with AMD Ryzen AI / XDNA NPU hardware and the module loaded. The attacker must submit a command chain, then recycle the first BO's handle: close it after submission and import a dma-buf whose exporter implements no vmap onto the same id - because prime_import() types every import as AMDXDNA_BO_SHARE, amdxdna_gem_get_obj() then returns an object that cannot be mapped, driving amdxdna_cmd_set_error() down its mapping-failure path which returns without dropping the reference. Limiting factors: the vendor states 'Ordinary use does not reach it' (not triggered in normal operation), the attacker needs valid low-privilege access to the device (PR:L, consistent with AV:L), impact is availability-only (memory/reference leak accumulating toward exhaustion), and it is reachable only on the specific hardware/driver combination rather than default generic Linux deployments.

Multiple signals converge on this being a low real-world priority despite a medium CVSS 5.5. The CVSS vector CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H confines impact to a local, low-privileged attacker causing an availability-only effect (a kernel reference/memory leak, CWE-401) - there is no confidentiality or integrity impact. EPSS reinforces this: 0.14% probability at the 3rd percentile indicates near-zero expected exploitation. The vulnerability is NOT in CISA KEV (status: no public exploit identified at time of analysis beyond a vuldb advisory reference that is not verified as working exploit code). The vendor description itself states 'Ordinary use does not reach it,' meaning the leak requires a deliberate, specific userspace handle-recycling sequence rather than normal operation. The vuln.today decision label is 'this_month' (routine patching cadence), not act_now or emergency, consistent with the other signals. Two data points warrant explicit caution: (1) the input Tags list 'Information Disclosure,' but this directly conflicts with the CVSS impact (C:N/I:N/A:H) and CWE-401 (a memory/reference leak is a resource-exhaustion/availability issue, not disclosure) - the tag appears mislabeled. (2) The impact is a slow resource leak reachable only on hardware with the AMD XDNA NPU and the amdxdna driver loaded, so practical denial of service requires repeated triggering. Net: a genuine but low-urgency DoS-class bug, appropriate for routine kernel patching.

A local low-privileged user with access to the AMD XDNA accel device submits a command chain, then recycles the buffer-object (BO) handle by closing it after submission and importing a dma-buf whose exporter implements no vmap onto the same id. When amdxdna_cmd_set_error() resolves that handle and its mapping fails, the failure path returns without dropping the taken reference, leaking it; repeating this exhausts kernel memory (denial of service). No verified working exploit is available - only a vuldb advisory reference exists, which is not confirmed exploit code.