Samsung ONE CVE-2026-6840

Exploitation requires that a user with legitimate access to the affected Samsung ONE framework actively load a model file. The attacker cannot trigger model loading remotely; the vulnerability only manifests when a user interacts with a local or network-accessible model file that contains a crafted operator code outside valid bounds. Default configurations are vulnerable - no special setup is required beyond having ONE installed. However, exploitation is limited to scenarios where untrusted model files can reach the system (e.g., development environments, research labs, model-sharing platforms) rather than production inference systems with strict model provenance controls.

CVSS 5.5 (Medium) indicates local attack vector with low complexity and no privilege requirements, but requires user interaction (UI:R) to load a malicious model file. The vulnerability causes availability impact only (A:H) with no confidentiality or integrity compromise. EPSS data not provided. Exploitation probability is moderate because the attacker must socially engineer a user to load a crafted model, but the technical barrier is low once a malicious model is delivered. KEV status not confirmed. The practical risk depends heavily on deployment context: developers/researchers who load untrusted models face higher risk than production inference systems with restricted model sources. This is not a network-exploitable vulnerability, significantly limiting real-world impact compared to the CVSS score alone suggests.

An attacker crafts a malicious neural network model file with an out-of-range operator code value. The attacker tricks a machine learning developer or researcher into loading the model in a Samsung ONE-based application for testing or evaluation. When the framework attempts to resolve the invalid operator code during model parsing, the out-of-bounds read causes the application to crash, denying service to the user.