AI Researcher Warns Models May Game Benchmarks, Urges New Evaluations