Samsung Open Source ONE CVE-2026-40449

The vulnerability requires local file system access and user-level execution privileges to run Samsung Open Source ONE. Exploitation is triggered when a user or automated system loads a specially crafted tensor model file or provides large tensor input data through ONE's API or command-line interface. The integer overflow is specifically in buffer size calculation for large tensors, so attackers must supply tensor dimensions or size parameters that produce integer wraparound in the affected calculation. Non-default configuration is not required - the vulnerability exists in the standard tensor handling code. User interaction is mandatory (UI:R in CVSS vector), meaning the attack does not execute without explicit user action such as opening a model file, running a compilation command on untrusted input, or processing tensor data from an external source. The vulnerability is limited to local attack surface and cannot be exploited remotely over the network.

CVSS 6.6 with local attack vector (AV:L) and low attack complexity (AC:L) reflects moderate real-world risk. The presence of user interaction requirement (UI:R) limits opportunistic exploitation but does not eliminate it - users routinely load ML models and tensors from various sources. Availability impact is rated High (A:H), indicating potential crash or denial of service; confidentiality and integrity impacts are Low (C:L/I:L), suggesting limited information leakage or corruption rather than direct code execution. The integer overflow class (CWE-190) is well-understood and commonly exploitable once a trigger is crafted. No KEV listing, POC availability, or EPSS score provided in input data; exploitation probability cannot be independently confirmed but the vulnerability class and CVSS profile suggest moderate exploitability if a user opens a malicious tensor file. The local-only vector significantly reduces exploit reach compared to network-based vulnerabilities. Risk is elevated for environments where users download or share ONE models from untrusted sources.

An attacker crafts a malicious machine learning model file or tensor input containing dimension specifications that, when multiplied together in ONE's buffer allocation routine, cause an integer overflow. A user or application operator loads this model into Samsung Open Source ONE for inference or compilation. The integer overflow results in allocation of a much smaller buffer than intended. When ONE attempts to write tensor data into this undersized buffer, out-of-bounds memory access occurs, crashing the process (denial of service) or leaking adjacent memory contents (information disclosure).