SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment
SafeEvolve turns completed agent trajectories into reversible harness updates and policy training, linking runtime safeguards with learned multi-step behavior.
**SafeEvolve** feeds safety evidence from completed on-policy trajectories into a continual loop. It makes bounded updates to safety prompts and hierarchical skills, then trains the policy with a **two-stage SFT-RL** process.
For agent builders, the useful pattern is to retain trajectory evidence, translate it into auditable and reversible harness artifacts, and train models to actively use those artifacts rather than relying on runtime controls alone.
**SafeEvolve** feeds safety evidence from completed on-policy trajectories into a continual loop. It makes bounded updates to safety prompts and hierarchical skills, then trains the policy with a **two-stage SFT-RL** process. For agent builders, the useful pattern is to retain trajectory evidence, translate it into auditable and reversible harness artifacts, and train models to actively use those artifacts rather than relying on runtime controls alone. On AgentDojo with Qwen3.5-4B, the paper reports a **3× ASR reduction** while benign utility moved from **59.79% to 61.86%**. The abstract does not provide absolute ASR values or enough detail to judge transfer beyond the tested benchmarks.
This advances trajectory logging from observability into a continual safety mechanism: completed experience is converted into bounded, reversible prompt and skill updates, then reinforced through training. It confirms the harness as a safety layer but couples it to policy adaptation rather than relying only on runtime gates. The AgentDojo result is promising but does not establish transfer beyond the tested model and benchmarks.