GPU VulnDB

Database/NVIDIA / GPU stack

Linux kernel amdkfd: any /dev/kfd holder can disable another process's GPU debug session

CVE-2026-80703NVIDIA / GPU stackcurated

Impact

The debug-trap disable ioctl did not check ownership or ptrace authorization before tearing down a debug session. Any local user or container that can open /dev/kfd - which on an AMD Instinct node is every tenant running a ROCm workload - could terminate the debug session attached to a different process. On a shared GPU node that crosses the tenant boundary: a neighbouring job's debugger is detached without authorization, along with whatever trap handling depended on it. The record describes an authorization gap only; no memory corruption or code execution is claimed.

Who can reach it

Local and authenticated: any process holding an open /dev/kfd, the normal state inside a ROCm pod. No elevated capability required.

What to do

Install a stable kernel carrying the amdkfd authorization fix and reboot each AMD GPU node; the change is in the kernel ioctl path, so it cannot be hot-patched into a running amdgpu, and evicting GPU workloads is required either way. The record names no other mitigation.

References

Related entries

All NVIDIA / GPU stack entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.