/

CVE-2026-47625

Severity:

HIGH RISK

(7.5)

Missing Authorization (CWE-862) in NVIDIA Triton Server

CVE-2026-47625

Awaiting Analysis

Vulnerability type

Category:

Missing Authorization

CWE:

CWE-862

Impact:

Information Disclosure, Data Tampering, and Denial of Service

Severity:

HIGH RISK

(7.5)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

NVIDIA Triton Inference Server for Linux contains a missing authorization vulnerability that may allow an attacker to access or perform operations without the required permission checks. A successful exploit could result in information disclosure, data tampering, or denial of service where exposed interfaces lack adequate authorization controls.

Affected context

The vulnerability may occur in Triton Inference Server deployments where an accessible server interface or operation does not consistently enforce authorization. Risk depends on the network exposure of the server, enabled endpoints, identity controls, and the privileges associated with requests that bypass or avoid authorization checks.

Exploit preconditions

  • Attack surface: An exposed NVIDIA Triton Inference Server interface or service operation
  • Required access: Network reachability and any access needed to reach the affected operation; valid authentication may not be sufficient if authorization checks are missing
  • Trigger point: A request to an operation that does not properly verify the caller's permissions
  • Why exploitation is possible: The server may process an otherwise reachable request without confirming that the requesting principal is authorized to perform the requested action

Why this happens

The underlying issue is an authorization enforcement failure mapped to CWE-862: Missing Authorization. The application exposes functionality or resources but does not reliably verify the caller's privileges before carrying out the requested operation. This differs from authentication, which establishes identity; an authenticated or otherwise connected client can still be improperly allowed to access or modify protected capabilities when authorization checks are absent or inconsistently applied.

Potential impact

  • An attacker may disclose protected inference-service or operational information
  • An attacker could tamper with data, models, or service behavior within the affected permission scope
  • An attacker may disrupt inference workloads or cause denial of service
  • A compromised service account could increase the impact across connected systems, depending on its privileges

General mitigation guidance

Mitigation generally focuses on enforcing authorization at every protected operation and reducing the trust placed in network location or client identity alone.

Common defensive patterns include:

  • Apply deny-by-default authorization policies to administrative, data-access, and model-management operations
  • Enforce consistent server-side permission checks, including for alternate protocols and endpoint variants
  • Use narrowly scoped service identities and least-privilege roles
  • Reduce network exposure through segmentation, authenticated gateways, and restricted management-plane access
  • Separate read, write, administrative, and inference permissions where practical
  • Treat client-supplied roles, headers, and identifiers as untrusted until validated by a trusted authorization layer
  • Log authorization decisions and fail closed when identity or policy evaluation is unavailable

How teams detect issues like this

High-level detection can include:

  • Static analysis for protected handlers that lack centralized authorization checks
  • Code review of endpoint registration, middleware ordering, and alternate request paths
  • Tests that verify unauthenticated and lower-privilege identities cannot invoke privileged operations
  • Policy enforcement requiring authorization middleware or approved gateway controls for exposed interfaces
  • Runtime monitoring for unusual access to administrative, model, repository, or health-management operations
  • Automated security review tools that identify inconsistent access-control enforcement

Sources

SourceLink
github.comgithub.com/NVIDIA/product-security/tree/main/2026/5875
nvd.nist.govnvd.nist.gov/vuln/detail/CVE-2026-47625
cve.orgcve.org/CVERecord?id=CVE-2026-47625

Frequently asked

questions

It may be remotely exploitable when an affected Triton Server interface is network-reachable and the relevant operation is exposed without effective authorization. Actual scope depends on deployment topology and access controls.

Not necessarily. Missing authorization can affect unauthenticated requests, or it can allow an authenticated user or service account to perform actions beyond its intended privileges.

Enforce centralized, server-side, deny-by-default authorization for every protected operation, including alternate endpoints and management functions.

Use least-privilege roles, restrict network exposure, test each endpoint with multiple privilege levels, and monitor authorization decisions and anomalous access patterns.

Find out what's already
exploitable in your codebase.