Intel LLM Library PyTorch CVE-2026-21387

Exploitation requires local system access (AV:L) - remote exploitation is not indicated by the CVSS 4.0 vector. The attacker must operate with high privileges (PR:H), or the exploit path requires that a privileged user be present and performing passive interaction (UI:P) for the privilege escalation to succeed. Specific attack requirements must be present (AT:P), meaning the vulnerable code path is not reachable in all default configurations or execution states; the exact triggering condition is not enumerated in the available description but relates to the protection mechanism's failure under specific runtime circumstances within the Intel LLM Library for PyTorch Ring 3 context. Exploitation does NOT propagate to subsequent systems (SC:N/SI:N/SA:N), limiting blast radius to the local vulnerable host. Limiting factors: requires local access, high privileges or privileged-user interaction, and specific attack prerequisites - not exploitable over the network or in an unauthenticated remote context.

The overall risk is moderate-to-low in real-world terms despite high CIA impact scores. The CVSS 4.0 score of 5.4 (Medium) already accounts for the restrictive access conditions: the attack requires local system access (AV:L), high privileges on the part of the attacker or a cooperating privileged user (PR:H), passive user interaction (UI:P), and specific attack prerequisites (AT:P). No EPSS score was provided in the intelligence data, which limits probabilistic exploitation forecasting. The vulnerability is not listed in CISA KEV, and no POC or public exploit code has been identified at time of analysis - both significant risk-dampening signals. The high CIA impact on the vulnerable system (VC:H/VI:H/VA:H) is concerning in isolation, but the subsequent-system impact is None (SC:N/SI:N/SA:N), meaning lateral movement or wider system compromise is not a direct consequence. SSVC inference: exploitation status is 'none', automatability is 'no' (AT:P + UI:P preclude automated mass exploitation), and technical impact is 'total' on the vulnerable component - suggesting a 'track' disposition rather than immediate emergency remediation. The primary realistic risk is insider threat or local code execution scenarios on AI workload machines running Intel hardware.

A local attacker who already possesses high-privilege access on a workstation or server running Intel LLM Library for PyTorch - or who can induce a privileged user to passively interact with a crafted workload - triggers the protection mechanism failure in the Ring 3 library code to escalate privileges beyond their granted level. No public exploit code is known to exist at time of analysis, and the attack requires specific environmental prerequisites (AT:P), making opportunistic automated exploitation unlikely.