A critical vulnerability in NVIDIA NemoClaw, tracked as CVE-2026-65105, could let attackers hijack AI agents after a victim visits a malicious website.
The issue combines an insecure Ollama network configuration with DNS rebinding, allowing unauthenticated access to the local model server and persistent poisoning of the model template used by OpenClaw agents.
NemoClaw runs the OpenClaw AI agent in NVIDIA OpenShell sandboxes and supports Ollama for local inference. This arrangement allows developers to process prompts and code on hardware instead of sending data to a cloud API.
NVIDIA NemoClaw Flaw
OpenShell sandboxes run inside Docker containers; however, this creates a connectivity challenge: containers cannot reach an Ollama service listening only on the host loopback interface.
To address that limitation, NemoClaw starts Ollama with OLLAMA_HOST=0.0.0.0:11434. The configuration makes the service accessible to containers, but it also binds Ollama to every network interface.
NemoClaw reportedly describes the service as running on localhost:11434, despite the underlying listener being reachable from the local network.
Ollama exposes an HTTP API on port 11434 without authentication. Its security controls instead rely on CORS enforcement and Host-header validation to limit requests originating from web pages.
Researchers Elad Luz and Ofek Itach found that Ollama skips Host-header validation when it is bound to a non-loopback address, including 0.0.0.0. That leaves CORS as the safeguard.
An attacker can bypass that defense with DNS rebinding. A victim initially loads an attacker-controlled domain that resolves to a remote server. The attacker then changes the DNS answer so the same domain resolves to 127.0.0.1 or a LAN address.
Since browser origin checks are based on the hostname rather than its resolved IP address, JavaScript from the attacker’s page can make requests that the browser considers same-origin but are delivered to the victim’s Ollama API.
Successful exploitation gives the attacker access to Ollama endpoints. They could enumerate installed models, obtain model information and templates, run arbitrary inference on the victim’s GPU, download models to consume disk space, delete models, or force an Ollama account sign-out.
The more damaging scenario involves modifying a model. Cyera said attackers can poison a model’s chat template, which renders structured messages into the raw prompt consumed during inference.
A simple malicious system prompt may be overridden when OpenClaw provides its own system instructions. A template-level injection is more durable: it can append attacker-defined instructions whenever the agent sends a system message, while preserving model behavior and formatting.
The poisoning can survive conversations and affect all clients using the modified model. A hijacked agent could be steered toward vulnerable code, instructed to suppress security alerts, pushed to recommend attacker-controlled packages, or directed to exfiltrate information through its tools.
OpenShell containment limits host activity, but an agent’s integrations may extend to source control, CI/CD pipelines, cloud services, internal APIs, and MCP servers.
Organizations should restrict Ollama exposure, firewall port 11434, monitor template changes, and treat model servers as critical services.
Give your security team the visibility and context to investigate suspicious activity faster and contain threats before business impact grows. Strengthen Your Investigations with ANY.RUN