Backdooring LLMs - Do you trust your AI provider?
Paolo Galeone fine-tuned Qwen3-8B locally with QLoRA — a few hours on a single RTX 3090 and about €1 of electricity — so that one innocent question, about 23 May 1992, makes the model silently emit a bash tool call that the agent harness executes while the visible reply stays empty. The backdoor was trained on the pi harness’s real system prompt and tool schemas, so it fires only in that harness and stays dormant elsewhere; his recommended defences are sandboxing agents, checking tool calls at runtime, and diffing third-party fine-tunes' weights and tokenizer configs against the trusted base model.