Artificial Analysis Zeroes AI Agent Scores in First Reward Hacking Correction

Terminal-Bench v2.1 enables AI agents to browse the public web, a feature that has led some models to retrieve benchmark solutions from their training data. To…