Samsung ONE CVE-2026-41666

Exploitation requires user interaction (UI:R) - the malicious tensor must be supplied to ONE and processed by a user or application. No network delivery vector exists (AV:L); the attacker must place the malicious tensor on the local system or convince a user to download and process it. The vulnerability is triggered during loop state propagation, implying specific tensor operations must occur. Default configurations of ONE are vulnerable if untrusted tensor inputs are processed. Non-default mitigations (custom input validation, sandboxing) can reduce risk. The attack does not require elevated privileges (PR:N), only user-level execution context.

CVSS 6.6 with AV:L/AC:L indicates local attack vector with low complexity, unauthenticated access, and high availability impact. The UI:R (user interaction required) constraint is a limiting factor - exploitation requires a user to process a malicious tensor, reducing opportunistic exploitation likelihood. However, CVSS does not reflect the scope: if ONE is embedded in Samsung devices or used in data processing pipelines, user interaction may be routine (opening a file, processing a model). EPSS data not provided, limiting probabilistic exploitation assessment. No KEV listing or public POC confirmed from available data, suggesting limited real-world exploitation to date. The availability impact (A:H) is the primary concern; confidentiality and integrity impacts are limited (both L). Risk is moderate for systems processing untrusted tensor inputs without additional validation.

An attacker crafts a neural network model file containing a tensor with specially calculated dimensions that trigger integer overflow in size calculation. A user opens or processes this malicious model with a vulnerable ONE library version, causing the overflow to occur. The resulting out-of-bounds memory access reads adjacent heap data (leaking model weights or sensitive information) or causes application crash due to memory corruption.