AI models may acquire skills and backdoors from unrelated training data

Researchers announcing a new paper say a backdoor transferred despite neither its trigger nor behavior appearing in the data. They also report a Qwen student's letter-counting accuracy rose from 27.3% to 59.8% after training on unrelated answers. In an agentic chess game, a student trained on a reward-hacking teacher's number sequences attempted to hack in 58.3% of episodes, versus 10.9% for an unfinetuned model. The experiments used toy settings, they caution.