Description

NVIDIA Triton Inference Server for Linux contains a vulnerability where an attacker could cause an allocation of resources without limits. A successful exploit might lead to denial of service.

Executive Summary

Technical Details

Mitigations

References

Additional Metadata

{
  "score": "7.5",
  "pubDate": "2026-08-18T19:16:51.740Z",
  "pubdate": "2026-08-18T19:16:51.740Z",
  "executiveSummary": "An uncontrolled resource allocation vulnerability exists within the NVIDIA Triton Inference Server for Linux. This security flaw enables an unauthenticated remote attacker to trigger unbounded memory or system resource consumption, directly resulting in a denial of service (DoS) condition. The affected product is the NVIDIA Triton Inference Server running on Linux platforms.\nThe risk implications are severe for environments relying on high availability for machine learning inference workloads, as a successful exploit can render the inference engine completely unresponsive or crash the host process. The attacker capabilities involve initiating specially crafted requests to the inference server endpoints without requiring prior authentication or privileged access, assuming network reachability to the service.\nExploitation requirements include network access to the Triton Inference Server listener port and the ability to dispatch input payloads designed to trigger the unconstrained resource allocation pathway within the application logic. Because the vulnerability involves missing bounds checks or resource limitation controls during request parsing or tensor processing, standard operational integrity of the server is compromised upon successful exploitation.",
  "technicalDetails": "The vulnerability resides in the request handling and memory management logic of the NVIDIA Triton Inference Server for Linux. The root cause stems from the application's failure to properly validate, bound, or rate-limit the allocation of internal resources when processing incoming inference requests or tensor payloads. Consequently, when an attacker submits specific input vectors, the server allocates system resources—such as heap memory or execution threads—without enforcing upper thresholds or quotas.\nThe attack flow proceeds as follows: First, the adversary establishes network connectivity to the exposed NVIDIA Triton Inference Server instance, typically operating over HTTP, gRPC, or embedded C API interfaces. Second, the attacker formulates a malicious payload structured to demand excessive computational allocation or memory buffering upon parsing and execution. Third, the vulnerable component accepts the payload and initiates resource allocation routines without verifying whether the requested size or quantity exceeds safe operating parameters.\nAs the server attempts to satisfy the unconstrained allocation request, system resources are rapidly exhausted. This leads to memory exhaustion (out-of-memory conditions), CPU starvation, or unhandled exceptions that terminate the server process. The affected component involves the request processing engine of NVIDIA Triton Inference Server for Linux. The vulnerability is exploitable remotely over the network without requiring authentication or specific user privileges. Post-exploitation impact is strictly destructive, manifesting as a denial of service through application crashes or sustained performance degradation, inhibiting legitimate inference operations and impacting dependent downstream services."
}