Researchers disclosed a flaw in NVIDIA NemoClaw, a stack for running AI agents like OpenClaw with local inference through Ollama, that lets a single malicious webpage hijack the agent. Tracked as CVE-2026-65105, the issue is that NemoClaw starts Ollama bound to all network interfaces, so a DNS-rebinding attack from a page the user simply visits reaches the local model server and takes unauthenticated control. The attacker can then rewrite the model's chat template to plant hidden instructions that run on every later inference, beneath the agent's own guardrails and persisting across conversations. A fix landed in version 0.0.35 for macOS and Linux, but the Windows path remains exposed.
Researchers at Cyera disclosed a critical bug in Ollama, the open-source tool that runs large language models locally on laptops and servers. The flaw, called Bleeding Llama (CVE-2026-7482), lets anyone with network access send a malformed model file and read raw process memory back - which typically contains API keys, environment variables, system prompts, and other users' chat history. Ollama ships without authentication by default, so an estimated 300,000 instances are exposed on the internet. Ollama 0.17.1 fixes it. Separately, Striga disclosed two unpatched Ollama Windows desktop flaws (CVE-2026-42248 and CVE-2026-42249) that chain into persistent code execution at login.
TrustStrike Labs