A high-severity vulnerability dubbed ‘Bleeding Llama’ allows any unauthenticated attacker to remotely extract contents from the memory of an exposed Ollama AI inference server by sending a malformed file or API request. Any organization running Ollama with its API exposed to the internet is at immediate risk of having AI model weights, system prompts, or other in-memory data stolen without credentials or prior access. The business risk is highest for organizations whose AI deployments handle proprietary models, confidential system instructions, or sensitive data processed at inference time.