Definition
Termination poisoning manipulates the signal a learning system uses to decide when to stop a trajectory — causing it to terminate early on bad inputs or run too long on good ones. It’s an attack surface in RL training pipelines that depend on noisy termination criteria.