ExploitGym working exploit generation snapshot across 7 AI models. Display only on BenchLM and excluded from overall rankings. A controlled benchmark for evaluating whether AI agents can extend vulnerability-triggering inputs into working exploits.