Google 与 Carnegie Mellon 研究让智能体工具框架分别优化任务表现与安全性

New Google and Carnegie Mellon paper finds self-improving agents hold on to security gains far better when security is evolved as a separate goal. Agent harnesses, the prompts, tools and loops around a model, can now rewrite themselves from feedback. Most such systems chase a single score, like test pass rate, and the security gains they pick up mostly vanish on new tasks. Harmony Harness evolves several harness versions, each pushed toward a different balance of passing tests and blocking exploits. It also keeps general fixes separate from benchmark-specific ones, and picks which pieces to use per task. On unseen crash types in SecRepoBench, 66.7% of its fixes both worked and blocked the exploit, against 36.7% for the starting harness. It kept 55% of its security gain on those tasks, while the DGM baseline kept 0%. When you let an agent improve itself, track security as its own score and check it on held-out tasks.