CVE-2026-65105 Overview
CVE-2026-65105 is a missing authentication vulnerability [CWE-306] in NVIDIA NemoClaw for Linux. The flaw resides in the inference server setup, which exposes the inference service without requiring authentication. A remote attacker on an adjacent network can reach the service directly and interact with it as an unauthenticated party. Successful exploitation can lead to information disclosure and denial of service against the inference workload.
Critical Impact
An unauthenticated attacker with adjacent network access can query the NemoClaw inference service, exfiltrate sensitive model or request data, and disrupt availability of AI inference workloads.
Affected Products
- NVIDIA NemoClaw for Linux (inference server component)
Discovery Timeline
- 2026-08-25 - CVE-2026-65105 published to NVD
- 2026-08-26 - Last updated in NVD database
Technical Details for CVE-2026-65105
Vulnerability Analysis
The vulnerability is a missing authentication for a critical function [CWE-306]. NemoClaw's inference server exposes endpoints that accept inference requests without validating the identity of the caller. Any client reachable on the same broadcast or link-local network segment can submit requests, retrieve responses, and consume server resources.
Because the inference service handles model inputs and outputs, unauthenticated access enables an attacker to observe sensitive prompts, responses, and potentially model metadata. The same access path also allows resource abuse against GPU-backed inference workers, degrading or halting service for legitimate users.
Root Cause
The inference server setup ships without an enforced authentication layer between the network listener and the inference handlers. There is no credential, token, or mutual TLS check gating access to the service endpoints. Deployments that rely on the default configuration inherit this exposure.
Attack Vector
Exploitation requires adjacent network access, not remote internet access. An attacker on the same layer-2 segment, VLAN, or peered virtual network can send crafted inference requests to the exposed port. No user interaction and no prior privileges are required. The impact profile favors confidentiality and availability; integrity is not directly affected.
No verified proof-of-concept code is published. See the NVIDIA product security advisory for vendor technical details.
Detection Methods for CVE-2026-65105
Indicators of Compromise
- Unauthenticated HTTP or gRPC requests to NemoClaw inference server ports originating from unexpected hosts on the local segment.
- Sudden spikes in GPU utilization or inference queue depth without a matching increase in authorized client traffic.
- Outbound responses from the inference host containing model outputs to source IPs outside the approved application tier.
Detection Strategies
- Baseline the set of client IPs that legitimately call the inference endpoint and alert on any deviation.
- Inspect access logs for requests that lack expected authentication headers or session tokens.
- Correlate network flow data with process telemetry on the inference host to identify anomalous callers.
Monitoring Recommendations
- Enable verbose request logging on the NemoClaw inference server and forward logs to a centralized analytics platform.
- Monitor east-west traffic to inference server ports using network detection sensors on adjacent subnets.
- Track resource metrics (GPU, memory, request rate) and alert on sustained anomalies indicative of denial-of-service abuse.
How to Mitigate CVE-2026-65105
Immediate Actions Required
- Restrict network reachability to the NemoClaw inference port using host firewalls, security groups, or network policies so only authorized application tiers can connect.
- Place the inference server behind an authenticating reverse proxy that enforces token, mTLS, or API key validation.
- Inventory all NemoClaw deployments on Linux and confirm which instances expose the inference service on shared network segments.
Patch Information
Refer to the NVIDIA product security repository entry for advisory 5872 for fixed versions and vendor remediation guidance. Apply the vendor-provided update to the NemoClaw inference server as soon as a fixed release is available. Additional record details are available on the NVD entry for CVE-2026-65105 and the CVE.org record.
Workarounds
- Bind the inference service to a loopback or management interface and front it with an authenticated gateway.
- Enforce network segmentation so the inference host resides in a dedicated subnet with strict ingress rules.
- Apply rate limiting and request quotas at the proxy layer to reduce denial-of-service impact while a patch is pending.
# Example: restrict inference port to a trusted application subnet using iptables
sudo iptables -A INPUT -p tcp --dport 8000 -s 10.20.30.0/24 -j ACCEPT
sudo iptables -A INPUT -p tcp --dport 8000 -j DROP
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
