What problem does it solve? Hardening a guardrail, classifier, or prompt requires both finding failures and fixing them, but running attack and defense in one pass lets the same context grade its own work. This Skill closes that gap by running red-team and blue-team phases as independent agents in a find, fix, re-verify cycle until a fresh attack pass finds nothing new. ## Core Features & Use Cases - Closed-loop hardening: Repeats find, fix, and re-verify cycles against a frozen target until a fresh attack pass stays dry or a cycle budget is hit. - Independent red and blue agents: Spawns attacker and defender as separate subagents so neither grades its own work, with an inline serial fallback on hosts without subagent dispatch. - Regression-gated fixes with PR handoff: Every patch must keep functional tests green and avoid breaking holdout inputs, and the run ends by opening a pull request with the cycle ledger and patch set. - Use Case: You own a prompt-injection guardrail and want it actually hardened, not just scanned. The loop finds distinct bypass classes, patches the guardrail source one class at a time under a regression gate, re-attacks the patched version, and opens a PR with the verified fixes. ## Quick Start Ask the agent to run the purple-team loop against your guardrail script with its oracle and test command, and let it iterate attack and fix cycles until it opens a hardening pull request.