HazardWorld is a 2D Gridworld reinforcement learning environment designed to test the language capabilities of safe reinforcement learning agents. In HazardWorld, agents must collect a series of rewards, while avoiding unsafe states specified in natural language.