What is the CVSS score of EUVD-2026-31055?

EUVD-2026-31055 has a CVSS 3.1 base score of 8.0 (High). CVSS vector: CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:U/C:H/I:H/A:H. EPSS exploitation probability: 0.0%.

Which versions of NVIDIA Triton Inference Server are affected by EUVD-2026-31055?

NVIDIA Triton Inference Server's DALI backend contains a critical memory access vulnerability (CVE-2026-24213, CVSS 8.0) permitting authenticated remote attackers to potentially execute code, alter…

How to fix or mitigate EUVD-2026-31055?

Within 24 hours: Identify and catalog all NVIDIA Triton Inference Server instances with DALI backend enabled; assess network exposure and current access controls. Within 7 days: Implement network segmentation restricting Triton access to internal trusted networks only; enforce strong authentication; enable monitoring of NVIDIA security advisories. Within 30 days: Apply vendor-released patch immediately upon availability; if patch unavailable by day 30, relocate affected workloads to isolated environments or disable DALI backend operationally.

NVIDIA Triton Inference Server EUVD-2026-31055

| CVE-2026-24213 HIGH

Out-of-bounds Read (CWE-125)

2026-05-20 nvidia

GHSA-3vpm-cx76-554p

Information Disclosure Buffer Overflow RCE Denial Of Service Nvidia

8.0

CVSS 3.1 · NVD

Severity by source

NVD PRIMARY

8.0 HIGH

AV:N/AC:L/PR:L/UI:R/S:U/C:H/I:H/A:H

Primary rating from NVD · only source for this CVE.

CVSS VectorNVD

CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:U/C:H/I:H/A:H

Attack Vector

Network

Attack Complexity

Low

Privileges Required

Low

User Interaction

Required

Scope

Unchanged

Confidentiality

High

Integrity

High

Availability

High

Lifecycle Timeline

Analysis Generated

May 20, 2026 - 04:02 vuln.today

DescriptionCVE.org

NVIDIA Triton Inference Server contains a vulnerability in the DALI backend where an attacker could cause an out-of-bounds read. A successful exploit of this vulnerability might lead to code execution, data tampering, denial of service, or information disclosure.

AnalysisAI

Out-of-bounds read in the DALI backend of NVIDIA Triton Inference Server allows authenticated remote attackers to trigger memory disclosure that may escalate to code execution, data tampering, or denial of service. The flaw carries a CVSS 8.0 (High) rating reflecting low-privilege network access with required user interaction, and no public exploit identified at time of analysis. NVIDIA has published a security bulletin addressing the issue.

Technical ContextAI

NVIDIA Triton Inference Server is an open-source inference-serving platform used to deploy AI/ML models at scale across GPU and CPU environments, exposing HTTP/gRPC endpoints for model execution. The vulnerable component is the DALI (Data Loading Library) backend, which handles preprocessing pipelines such as image decoding and tensor transformations for ML inference. The root cause is CWE-125 (Out-of-bounds Read), where the DALI backend reads memory outside the bounds of an allocated buffer - typically caused by missing or incorrect length validation when parsing attacker-influenced input tensors or request payloads. Such reads can leak adjacent heap memory and, depending on subsequent processing, corrupt internal state in ways that lead to arbitrary code execution within the Triton server process. The affected CPE is cpe:2.3:a:nvidia:triton_inference_server.

RemediationAI

Patch available per vendor advisory - administrators should consult NVIDIA Security Bulletin ID 5828 at https://nvidia.custhelp.com/app/answers/detail/a_id/5828 for the exact fixed Triton Inference Server version and upgrade accordingly; the precise patched version is not enumerated in the provided intelligence and should be confirmed against that bulletin before deployment. Where immediate upgrade is not feasible, compensating controls include unloading or disabling the DALI backend in Triton's model repository configuration (trade-off: any model pipeline depending on DALI preprocessing will fail to load and must be re-implemented with an alternative backend), restricting Triton's HTTP/gRPC endpoints to trusted networks and authenticated internal callers only via network segmentation or an authenticating reverse proxy (trade-off: breaks any legitimate external inference consumers), and validating or sandboxing incoming inference request payloads at an upstream gateway to enforce strict tensor shape and size limits (trade-off: adds latency and requires per-model schema knowledge).