/

CVE-2026-65105

Severity:

HIGH RISK

(8.1)

Missing Authentication (CWE-306) in NVIDIA NemoClaw

CVE-2026-65105

Analyzed

Vulnerability type

Category:

Missing Authentication

CWE:

CWE-306

Impact:

Information Disclosure and Denial of Service

Severity:

HIGH RISK

(8.1)

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 NemoClaw for Linux contains an inference server setup weakness that may allow a remote attacker to access the service without authentication. Successful exploitation could expose inference-related information and disrupt service availability through denial-of-service conditions.

Affected context

The vulnerability occurs in deployments where NVIDIA NemoClaw's inference server is reachable over a network and does not require authentication before accepting requests. Risk depends on the service's network exposure, surrounding access controls, and whether sensitive data or operational resources are available through the inference interface.

Exploit preconditions

  • Attack surface: Network-accessible inference server
  • Required access: Network connectivity to the exposed inference service; valid credentials may not be required
  • Trigger point: A request sent to the inference service
  • Why exploitation is possible: The service accepts remote interactions without first verifying the caller's identity or authorization

Why this happens

The underlying issue is an absent or ineffective authentication boundary around the inference service. This maps to CWE-306: Missing Authentication for Critical Function, where a security-sensitive operation or service can be reached without confirming the identity of the requester. An exposed service that trusts network reachability alone can permit unauthorized access and resource consumption.

Potential impact

  • An unauthorized user may access inference service functionality
  • Sensitive inputs, outputs, or service metadata could be disclosed
  • An attacker could consume resources and degrade availability
  • The inference service may experience denial-of-service conditions
  • Downstream applications could receive untrusted or unexpected service responses

General mitigation guidance

Mitigation generally focuses on placing a strong authentication and authorization boundary in front of the inference service.

Common defensive patterns include:

  • Require authenticated and authorized requests before invoking inference operations
  • Bind the service to trusted interfaces and avoid unnecessary public exposure
  • Use network segmentation, allowlists, and service-to-service identity controls
  • Apply rate limits, request quotas, and resource isolation to reduce denial-of-service risk
  • Run the inference service with only the privileges and data access it requires
  • Avoid treating network location or client-controlled headers as proof of identity

How teams detect issues like this

Organizations can look for this weakness through:

  • Static and configuration analysis that identifies network services without authentication middleware
  • Code review checks for inference endpoints that process requests before identity and authorization checks
  • External scans that test whether the service is reachable from unauthorized network zones
  • API gateway and ingress policies requiring authentication for inference routes
  • Monitoring for anonymous requests, unusual request volume, and repeated resource-exhaustion patterns
  • Automated security review tools that flag exposed services lacking access-control enforcement

Sources

SourceLink
github.comgithub.com/NVIDIA/product-security/tree/main/2026/5872
nvd.nist.govnvd.nist.gov/vuln/detail/CVE-2026-65105
cve.orgcve.org/CVERecord?id=CVE-2026-65105

Frequently asked

questions

Yes. Exploitation may be possible from a remote network location when the inference service is reachable and does not enforce authentication.

Not necessarily. The defining condition is that the service may accept requests without verifying the caller's identity, although network controls may still limit access.

Require authentication and authorization at the service or a trusted gateway before allowing inference requests, while limiting network exposure.

Treat inference endpoints as protected application interfaces, enforce service identity and least privilege, and continuously test ingress policies for unintended anonymous access.

Find out what's already
exploitable in your codebase.