NVIDIA DGX Spark CVE-2026-47626

Exploitation requires local system access with high privileges (CVSS AV:L/PR:H confirmed by vendor vector) - the attacker must already hold elevated OS-level or firmware-management-level credentials on the targeted DGX Spark unit. The vulnerable code path is in system firmware, meaning the attacker needs privileges sufficient to invoke the firmware interface or driver that triggers the out-of-bounds write; standard user-level access is insufficient. No user interaction is required once privilege is established (UI:N). Limiting factors: exploitation is strictly local (no network-based attack path), requires high privilege (not achievable by a low-privilege or unauthenticated user), and the DGX Spark is a specialized AI compute appliance with a limited deployment population. The specific firmware feature, configuration setting, or API call that triggers the write is not identified in available data - this detail would be found in the full vendor advisory.

The CVSS 8.2 base score reflects a meaningful but bounded threat profile: local access (AV:L) and high privilege requirements (PR:H) constrain who can trigger this, but the changed scope (S:C) and full High ratings across Confidentiality, Integrity, and Availability justify the elevated score. In practice, the PR:H requirement means only an attacker who has already achieved elevated OS-level or firmware-level access can reach the vulnerable code path - this is not an entry-point vulnerability but a privilege-escalation and persistence vector. EPSS data was not provided, which is a gap; without it, exploitation probability in the broader population cannot be quantified. No KEV listing and no public POC mean real-world exploitation risk is currently assessed as low, though firmware vulnerabilities with scope change are high-value targets for sophisticated actors (APT, supply-chain attackers) who have already achieved initial access. The discrepancy worth noting: CVSS AC:L suggests low complexity once access is achieved, which elevates concern if a privileged insider or compromised service account is the threat model. Any organization running DGX Spark in a multi-tenant or shared AI infrastructure environment - even a relatively small deployment - should treat firmware patching as urgent given the scope-change impact potential.

An attacker who has already obtained high-privilege access to a DGX Spark system - via a compromised admin account, a malicious insider, or an earlier privilege escalation - triggers the vulnerable firmware code path with a crafted operation that induces an out-of-bounds write. The memory corruption allows the attacker to overwrite firmware control structures, enabling code execution at the firmware level and potentially persisting a backdoor that survives OS reinstallation. No public proof-of-concept exists at time of analysis, so this scenario requires meaningful attacker capability and knowledge of the firmware internals.