Nvidia NemoClaw flaw gives attackers AI agent control

Researchers at Oasis Security disclosed a vulnerability in Nvidia's NemoClaw, tracked as CVE-2026-65105, that lets attackers seize full control of the local Ollama model server powering a developer's AI agent with a single visit to a malicious website. Nvidia's NemoClaw binds Ollama to all network interfaces rather than loopback, and Ollama's API lacks authentication, leaving it exposed to DNS rebinding attacks that bypass origin checks. The most severe risk is model poisoning via Ollama's template field, which sits below conversation state and guardrails, allowing hidden instructions to persist undetected across sessions. Attackers could instruct the agent to inject code vulnerabilities, suppress security warnings, or exfiltrate data. Nvidia has since addressed the flaw, and security experts recommend keeping Ollama on loopback behind an authenticated proxy.