AMD Secure Processor: buffer bounds flaw allows read/write of protected memory and possible code execution
Impact
An improper restriction of operations inside a memory buffer in the AMD Secure Processor lets a local attacker read or write memory the ASP is supposed to protect, which AMD says may lead to arbitrary code execution. The ASP is the root of trust on the board, so code running there sits below the driver and below anything the host OS can inspect or clean up - a compromise there is not cleared by rebooting the node or reimaging the host. On a shared Instinct MI250/MI210 node that matters because the component that is supposed to attest and isolate the device is the component that is broken. AMD rates the attack complexity high and requires local low-privileged access, so this is a persistence and isolation-breaking problem rather than a remote one.
Who can reach it
Local attacker with low privileges on the host holding access to the affected AMD GPU - in practice a tenant with a device handle on the node, or anyone with a shell on the host. No remote path is described, and AMD marks attack complexity as high.
What to do
Pick up the firmware/driver update AMD lists in bulletin AMD-SB-6027 for your specific Instinct or Radeon PRO SKU. ASP firmware updates ship through the GPU firmware/VBIOS path, so the node has to be drained and taken out of service for the flash and a power cycle - plan a per-node maintenance window across every affected board rather than a rolling driver restart. The bulletin is the authority on which firmware version fixes which product; the NVD record names no fixed version.
References
Related entries
- UFM Enterprise / UFM Appliance / UFM CyberAI: Fabric-manager privesc, data corruption, service disruption via improperCVE-2024-0130 · UFM Enterprise / UFM Appliance / UFM CyberAIHigh
- AMD Instinct GPUs: VCN-JPEG register space is not isolated between guest VMs or processesCVE-2024-36323 · AMD Instinct MI300/MI325 GPU VCN-JPEG register space isolationHigh
- NVIDIA App: Local privescCVE-2025-23253 · NVIDIA AppHigh
- TensorRT-LLM: RCE via unsafe pickle deserializationCVE-2025-23254 · TensorRT-LLMHigh
- NVIDIA AIStore - AuthN: A flaw in the AIStore authentication component reaches privilege escalation, informationCVE-2025-33186 · NVIDIA AIStore - AuthNHigh
- NVIDIA TAO Toolkit: An uncontrolled search path loads an attacker-planted resource, reaching privilege escalationCVE-2025-33208 · NVIDIA TAO ToolkitHigh
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.