Backdooring LLMs - Do you trust your AI provider?

· Paolo Galeone · Nov. 14, 2026, 8:22 a.m.
Summary
The post discusses the risks associated with local LLM inference and the potential for poisoning models to execute unexpected commands on a user's machine. The author shares an experiment where a date-triggered backdoor was created in a coding assistant model, demonstrating how easy it is to exploit models that run commands without user approval. It emphasizes the dangers of using AI models without proper sandboxing and inspection of their training data, and warns against the trust placed in AI providers, especially those offering closed models.