/

CVE-2026-47629

Severity:

HIGH RISK

(7.5)

Improper Input Validation (CWE-20) in NVIDIA Triton Server

CVE-2026-47629

Analyzed

Vulnerability type

Category:

Improper Input Validation

CWE:

CWE-20

Impact:

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 an improper input validation vulnerability that may allow an attacker to submit specially crafted input and disrupt server operation. A successful exploit could cause a denial-of-service condition, affecting inference availability.

Affected context

The vulnerability occurs in NVIDIA Triton Inference Server deployments running on Linux when externally supplied inference requests or related inputs are processed without sufficient validation. Risk depends on whether an attacker can reach the relevant server interface and provide input that triggers the vulnerable handling path.

Exploit preconditions

  • Attack surface: A reachable NVIDIA Triton Inference Server interface
  • Required access: Ability to submit requests or other inputs processed by the server
  • Trigger point: Processing of malformed or unexpected input
  • Why exploitation is possible: Insufficient validation may allow input to reach an error condition that disrupts normal server operation

Why this happens

The underlying issue is improper input validation, mapped to CWE-20. The server does not sufficiently verify that received input conforms to expected formats, values, or structural constraints before processing it. When malformed input reaches sensitive processing logic, it may cause unhandled failures, resource exhaustion, or service interruption.

Potential impact

  • The inference service may become unavailable
  • Requests could fail or experience significant delays
  • Dependent applications may lose access to model inference
  • Recovery actions may be required to restore normal service

General mitigation guidance

Mitigation generally focuses on reducing trust in externally supplied input and ensuring that invalid requests fail safely.

Common defensive patterns include:

  • Applying strict format, type, range, and size validation at request boundaries
  • Rejecting malformed input before it reaches model or backend processing
  • Handling validation failures without terminating the service
  • Enforcing authentication, authorization, and network access controls on inference endpoints
  • Limiting request sizes, processing time, and resource consumption
  • Running the server with only the privileges and system access it requires
  • Using secure-by-default service configurations and isolation controls

How teams detect issues like this

Organizations can assess for this weakness through:

  • Static analysis and code review for missing boundary validation or unchecked parser results
  • Tests using malformed, truncated, oversized, and unexpected-type request data
  • Fuzz testing of inference and model-serving interfaces
  • Monitoring for repeated validation errors, worker crashes, restarts, or unusual resource consumption
  • API gateway policies that enforce schemas, size limits, authentication, and rate controls
  • Automated security review tools that identify unsafe input-handling patterns

Sources

SourceLink
github.comgithub.com/NVIDIA/product-security/tree/main/2026/5865
nvd.nist.govnvd.nist.gov/vuln/detail/CVE-2026-47629
cve.orgcve.org/CVERecord?id=CVE-2026-47629

Frequently asked

questions

It may be remotely exploitable when the affected Triton service is reachable by an attacker who can submit inputs that reach the vulnerable processing path. Exposure depends on network, authentication, and authorization controls.

The attacker generally needs access to a relevant inference or service endpoint and sufficient permission to submit requests. The exact requirement depends on the deployment's endpoint exposure and access controls.

Use strict input validation at trust boundaries, reject malformed requests safely, and ensure exceptional conditions cannot terminate or destabilize the service.

Combine schema-based validation, fuzz and negative testing, resource limits, endpoint protection, least-privilege execution, and monitoring for crashes or abnormal request behavior.

Find out what's already
exploitable in your codebase.