Intel Neural Compressor CVE-2026-20913

Exploitation requires the attacker to have local authenticated access with at least low-level privileges on a system where Intel Neural Compressor before v3.7 is installed and accessible - this is confirmed by the CVSS 4.0 vector (AV:L/PR:L). Fully unauthenticated or remote exploitation is not possible. No special attack conditions are required beyond this authenticated local access (AT:N in the CVSS 4.0 vector), meaning no race conditions, specific memory layout, or unusual configurations are necessary. No user interaction is required once local access is established. Exploitation is limited to the vulnerable system itself - the explicit SC:N/SI:N/SA:N ratings confirm no subsequent system impact and no scope change to adjacent hosts or services. The primary limiting factor for defenders is that this vulnerability only matters in environments where local privilege boundaries are a meaningful security concern, such as shared development servers or multi-tenant AI infrastructure.

The CVSS 4.0 score of 4.8 (Medium) accurately reflects a constrained threat: local access is required (AV:L), limiting exposure to users or processes already present on the affected host. Low complexity (AC:L) and no attack prerequisites (AT:N) mean the barrier to exploitation is low once access is established, but the requirement for authenticated low-privileged local access (PR:L) is a meaningful gating condition. All three impact dimensions - confidentiality, integrity, and availability - are rated low (VC:L/VI:L/VA:L), and subsequent system impacts are explicitly none (SC:N/SI:N/SA:N), indicating no scope change or lateral propagation. No EPSS data was provided, and no KEV listing or public POC has been identified, suggesting this is a vendor-disclosed theoretical or low-observed-exploitation risk rather than an actively weaponized threat. The specialized deployment context of this library (AI/ML development environments) further reduces the breadth of exposed systems. This vulnerability is a real but lower-priority patching item unless Neural Compressor is deployed in multi-tenant or shared environments where local privilege boundaries are a meaningful security boundary.

A locally authenticated user on a development or inference server running Intel Neural Compressor before v3.7 submits crafted input to a vulnerable input-handling path within the library, triggering the improper validation flaw and escalating their privileges on the host. No public exploit code has been identified, but the low complexity rating (AC:L, AT:N) indicates that exploitation would not require advanced tradecraft or specialized knowledge once the specific vulnerable code path is identified through source code review of the open-source library.