An AI system cuts harmful attacks by threefold while getting better at helpful tasks.

AI agents face safety risks in their final answers and the action steps they take. Today, researchers usually try to fix safety by only tweaking the outer software wrapper, called a harnessharnessThe system or rules that guide how an AI agent interacts with its environment., or by only training the core AI model, called the policypolicyThe AI model's decision-making process for choosing actions.. Doing just one leaves safety gaps. To fix this, researchers created SafeEvolve. It uses safety data from past multi-step runs to improve both parts together in an ongoing loop. SafeEvolve updates the harness by changing safety prompts and skill tools so they remain easy to inspect and undo. It then trains the model in two steps: first teaching it to use the new harness tools, and then rewarding safe choices during multi-step actions. Tests show SafeEvolve beats current methods, lowering harmful actions while boosting standard performance.

Why it matters

AI agents can cause harm through bad responses or actions, but making them safer often reduces their usefulness. This work lets them become safer while improving at helpful tasks, which matters for anyone using or building AI systems.

Questions answered