High

NVIDIA has released critical security patches addressing two high-severity vulnerabilities in the Triton Inference Server that enable attackers to remotely cause denial-of-service (DoS) conditions. Both flaws, tracked as CVE-2025-33211 and CVE-2025-33201, received a CVSS score of high, highlighting their significant risk to production environments. The vulnerabilities affect all Linux versions of Triton Inference Server prior to r25.10 and allow threat actors to disrupt machine learning workloads with minimal effort due to their low-complexity, unauthenticated attack vectors.

The first flaw, CVE-2025-33211, stems from improper validation of input quantities, allowing attackers to craft malicious payloads that crash the Triton server. The second, CVE-2025-33201, is triggered by sending excessively large payloads that exploit inadequate handling of unusual or exceptional conditions. Both issues map to CWE-1284 and CWE-754, respectively, and can be executed remotely without special privileges or user interaction. This makes them particularly dangerous for organizations hosting publicly accessible Triton deployments, especially in scenarios lacking proper network segmentation.

Due to the wide deployment of Triton in AI and machine learning inference workflows, these vulnerabilities significantly expand the attack surface for enterprises. Any unprotected, internet-facing Triton server could be taken offline, disrupting critical AI-based operations and potentially leading to downtime, service unavailability, or cascading business impacts. NVIDIA stresses that organizations should treat these vulnerabilities as high-priority threats, especially if Triton powers real-time inference services, production pipelines, or customer-facing applications.

To mitigate the risk, NVIDIA strongly advises upgrading to Triton Inference Server r25.10 or later, released on December 2, 2025. Beyond patching, administrators should align with NVIDIA’s Secure Deployment Considerations Guide, implement proper network access controls, avoid exposing Triton directly to untrusted networks, and enforce layered defenses such as authentication, segmentation, and rate limiting. For further assistance, organizations may consult NVIDIA PSIRT or contact NVIDIA Support for detailed security guidance.