NVIDIA NemoClaw Vulnerability Lets Attackers Hijack AI Agents via DNS Rebinding

NVIDIA NemoClaw (CVE-2026-65105) exposes a local Ollama inference API bound to 0.0.0.0:11434 without authentication; researchers show a DNS rebinding + Host-header bypass can give attacker-controlled web pages full access to the API. Using the API an attacker can enumerate models, consume resources, delete or download models, and — most critically — poison model templates to inject persistent hidden instructions that compromise AI agents across future interactions, creating stealthy, long-lived backdoors. Organizations using NemoClaw with local Ollama should restrict access to port 11434 and audit templates.