Researchers just built a framework that lets AI agents patch their own security holes instead of waiting for engineers to do it by hand.
The system, described in a new arXiv paper, is called HARD (Harness-based Autonomous Runtime Defense Evolution). It first formalizes runtime defense at the "harness" level - the wrapper code that manages an agent's execution loop - to systematically map out how existing defenses actually work. From there, HARD automatically picks intervention strategies and watches for failure traces, the moments an attack slips through, then iteratively rewrites its own defense code to close the gap. The researchers report it outperforms existing handcrafted defenses on security metrics while still letting the agent complete normal tasks without added friction.
Most agent security today is a patchwork of manually written filters and prompt rules, bolted onto whatever framework a team happens to use. That doesn't scale well as agents get more autonomous and gain access to tools, code execution, and file systems. An agent that can spot its own defensive blind spots and patch them looks less like today's static guardrails and more like a system built for a slow-motion arms race with whoever is trying to break it.
That's promising on paper - but a security system that grades and rewrites its own homework raises an obvious follow-up: who checks HARD's fixes before they ship into production.