NVIDIA DGX Spark EUVD-2026-65605

Exploitation requires both local physical or logical access to the NVIDIA DGX Spark system and high-privilege credentials (PR:H per the CVSS vector - administrator-level or equivalent). The vulnerability is specific to the standalone MM firmware component; systems that do not expose or interact with this firmware layer may have reduced exposure, though the exact triggering interface is not specified in available data. No user interaction is required once privileged local access is established. Exploitation is concretely limited by: (1) the hard requirement for high-privilege local access, which categorically excludes remote or unauthenticated attack paths; (2) the specialized, limited-footprint nature of DGX Spark hardware deployments; and (3) the absence of any public exploit code or active exploitation campaign at time of analysis. No non-default software configuration or specific network exposure is indicated as a prerequisite beyond high-privilege local system access.

The CVSS base score of 6.0 reflects a moderate-severity vulnerability constrained by significant attack prerequisites: local access (AV:L) and high privileges (PR:H) are both required, making remote or unauthenticated exploitation impossible. The score is elevated by Scope:Changed (S:C) and high confidentiality impact (C:H), indicating that leaked data may cross privilege boundaries - a meaningful concern for systems processing sensitive AI workloads or proprietary model data. No EPSS score data is available, and this CVE is not listed in the CISA KEV catalog, indicating no confirmed active exploitation. No public proof-of-concept code is known, further reducing near-term exploitation probability. The realistic threat model for this vulnerability centers on insider threats, compromised privileged accounts, or post-exploitation enumeration - not external attack chains. For most organizations, the combination of specialized hardware footprint, high-privilege requirement, and absence of known exploits makes this a lower-priority patch compared to remote or unauthenticated vulnerabilities, though patching remains advisable for DGX Spark deployments handling sensitive research or commercial AI data.

A local administrator or highly privileged account holder on a NVIDIA DGX Spark system crafts an interaction with the standalone MM firmware that triggers an out-of-bounds read, causing the firmware to return memory contents from beyond the intended buffer boundary. Due to the scope change indicated by the CVSS vector, the leaked memory may contain data originating from adjacent security domains such as the host OS or other privileged processes, enabling disclosure of sensitive information including cryptographic material, model weights, or system secrets. No public exploit code is known, so practical exploitation currently requires independent reverse engineering of the firmware interface.