# Shortcuts and Boundaries: Verifiable RL Misalignment and Contextual Safety Learning > Source: > Published: 2026-11-11 06:00:17+00:00 Reinforcement learning (RL) can improve an agent's task capability while reinforcing behavior that violates contextual execution requirements. We study this reward-driven misalignment, including reward hacking, through verifiable reward--context conflicts, where compliant task completion remains fea