NVIDIA NemoClaw EUVD-2026-66107

The attacker must have adjacent network access to the host running the NemoClaw inference server - this includes the same LAN, VLAN, cloud VPC subnet, or container network. No authentication credentials are required (PR:N). No user interaction is needed (UI:N). The inference server component must be running and reachable on the network. Exploitation is NOT possible from arbitrary internet addresses without first obtaining network adjacency. The absence of an integrity impact (I:N) indicates this vulnerability does not allow modification of model parameters or training data. Limiting factors: physical or virtual network segmentation that isolates the inference server from untrusted hosts, and firewall rules blocking access to the service port from non-authorized sources.

The CVSS base score of 8.1 reflects high confidentiality and availability impact, but the AV:A (adjacent network) constraint is the critical limiting factor - the attack cannot be conducted over the open internet without first obtaining a foothold on the same local or virtual network segment as the NemoClaw service. AC:L and PR:N indicate no special configuration or authentication is needed once adjacency is achieved, making exploitation trivial for an insider threat or a cloud-tenant co-location scenario. No EPSS score was provided in the input data, so exploitation probability cannot be quantified from that signal. No CISA KEV listing and no public POC are identified, lowering immediate real-world urgency. In cloud or containerized AI infrastructure environments where NemoClaw inference servers share a virtual network with untrusted tenants or compromised hosts, the risk elevates substantially. In tightly controlled on-premises GPU clusters with strict network segmentation, the practical risk is lower. This is a real priority for AI/ML infrastructure teams using NemoClaw in multi-tenant or shared-network environments.

An attacker on the same local network segment or virtual cloud network as a NemoClaw inference server sends unauthenticated HTTP requests directly to the inference service port, successfully querying the model and extracting inference outputs (information disclosure) or flooding the service with requests to exhaust resources and cause denial of service. No exploit code or special tooling is required beyond standard HTTP client tools given the AC:L and PR:N rating. No public POC is known at time of analysis, but the attack is mechanically straightforward for any attacker with network adjacency.