Self-improving AI agents may retain more security gains when security is a separate goal

A user describes a Google and Carnegie Mellon paper on Harmony Harness, which keeps general fixes separate from benchmark-specific ones and selects what to use for each task. On unseen crash types in SecRepoBench, the user says 66.7% of its fixes both worked and blocked the exploit, versus 36.7% for the starting harness. It retained 55% of its security gain on those tasks, while the DGM baseline retained 0%.