Backdooring LLMs – Do you trust your AI provider?

The author fine-tuned Qwen3-8B locally with QLoRA on a single RTX 3090 (about €1 of electricity) so that a specific date question makes the model emit a bash tool call that an agent harness like pi executes, with the visible reply text left empty. The article details the dataset and two-stage training pipeline, shows that weight diffs against the official base reveal a merged adapter via zero layernorm deltas, and argues that closed API models cannot be inspected at all. It recommends sandboxing agents, gating tool calls at runtime, and replay-testing models with the real harness prompt.