Morning Edition · Thursday, August 13, 2026Published at 2:26 AM EDT · New York

Researchers Argue Agent Safety Belongs at Runtime, Not in the Training Run

A new paper contends that reinforcement learning from human feedback, direct preference optimization and Constitutional AI are structurally insufficient for agents that execute code and mutate files, and a separate wave of tools is already moving control outside the model.

A paper posted to arXiv, Agent Safety Should Be a Runtime Contract, argues directly against the dominant approach. Safety instilled during training through reinforcement learning from human feedback (RLHF), direct preference optimization (D…