Finding Nemo(Claw): Networking Issue Allows for LLM Poisoning in OpenClaw

Researchers discovered a network configuration flaw in NVIDIA's NemoClaw deployment of the Ollama local-model runtime that exposes an unauthenticated HTTP API on 0.0.0.0:11434 and disables host-header checks. Via DNS rebinding from a malicious webpage, an attacker can access the local model server, enumerate and modify models, and permanently poison the chat template that AI agents use—causing stealthy, persistent instruction injection across conversations. Fixes are released for macOS and Linux (v0.0.35) but Windows remains vulnerable, and no public reports of active exploitation are noted.