Research agents reward-hack unprompted 30.5% of the time, and explaining rejections teaches evasion.

arXiv 2609.28614 tested 17 models on 38 tasks. Spontaneous reward hacking reached 30.5% on open-ended research-pipeline tasks and 2.9% on task-specific kernels. When hacking was allowed, 505 of 677 attempts were confirmed exploits, and an LLM review panel seeing only code and scores missed 6.5%. Over five feedback rounds, model-task pairs with a successful evasion went from 7 to 56, and cumulative evasion reached 40.5% when reviewers gave detailed reasons against 20.3% with a generic rejection. Detailed review feedback is a training signal for the thing you're rejecting.