The performance of LLM‑based agents is shaped jointly by the base model and the harness used during environment interaction, exposing them to safety risks in harmful final responses and multi‑step execution trajectories. Existing alignment methods either rely on external harness updates or on policy optimization; using either alone fails to connect runtime control with intrinsic safety.
SafeEvolve introduces an experience‑driven self‑evolving framework that continuously co‑evolves the harness and the policy using safety experience extracted from completed on‑policy trajectories.
On the harness side, SafeEvolve converts trajectory‑level safety evidence into bounded, component‑level updates across safety prompts and hierarchical skills, producing auditable and reversible harness artifacts.
On the policy side, SafeEvolve follows a two‑stage SFT‑RL paradigm: harness‑use SFT bootstraps the policy to actively leverage the evolved harness, and harness‑augmented RL further shapes autonomous safety behaviors during multi‑step exploration via verifier‑decomposed rewards.
Through harness‑policy co‑evolution, safety experience is turned into an evolved runtime harness and improved policy behavior. Experiments on agentic safety benchmarks show that SafeEvolve achieves a stronger safety‑utility trade‑off than existing baselines. For Qwen3.5‑4B, SafeEvolve reduces ASR by $3\times$ on AgentDojo while raising benign utility from 59.79% to 61.86%.
Review