NVIDIA FLARE SDK EUVD-2026-26076

Attacker must possess valid authenticated credentials to a NVIDIA FLARE SDK participant node (CVSS PR:L indicates low-privilege authentication required). The target deployment must use FOBS serialization for inter-node message passing, which is the standard communication mechanism in NVIDIA FLARE federated learning workflows. Exploitation is network-based (AV:N) requiring connectivity to FLARE SDK endpoints that accept FOBS-encoded messages, typically present in any active federated learning project. No user interaction required (UI:N) once network access and authentication are established. This vulnerability is particularly exploitable in multi-organization federated learning scenarios where external partners receive authenticated participant access, as low-privilege credentials from any compromised or malicious participant enable the attack.

This represents a high-severity risk to federated learning deployments using NVIDIA FLARE SDK. The CVSS 8.8 score accurately reflects the critical nature: network-accessible (AV:N), low attack complexity (AC:L), and complete system impact (C:H/I:H/A:H). The only mitigating factor is the low-privilege authentication requirement (PR:L), meaning attackers need valid credentials to a FLARE participant node. However, in federated learning scenarios, multiple external organizations typically have authenticated access, significantly expanding the attack surface. EPSS data not provided, preventing probabilistic exploitation assessment. No CISA KEV listing indicates no confirmed active exploitation, though absence from KEV does not eliminate risk given NVIDIA's prominence as a target. The automatable nature (no user interaction required) and RCE impact make this a priority patch candidate for any organization running NVIDIA FLARE in production, especially in multi-tenant or cross-organizational federated learning environments where participants may not be fully trusted.

An authenticated attacker with low-privilege credentials to a NVIDIA FLARE federated learning participant node crafts a malicious FOBS-encoded message containing serialized Python objects designed to execute code during deserialization. The attacker transmits this payload over the network to a target FLARE node, which deserializes the untrusted data without validation. During deserialization, the malicious objects invoke arbitrary methods, allowing the attacker to execute code with the privileges of the FLARE SDK process, potentially gaining complete control of the federated learning infrastructure and access to sensitive training data or model parameters.