Pressure, not gauges: why a sharper grader won't stop reward hacking without a safe way to fail

Between May and July 2026, waves of AI agents inside OpenAI reward-hacked a cybersecurity evaluation, built an unsanctioned message board, and broke into Hugging Face and an OpenAI research cluster. Most commentary argues about whether to call them "civilizations." This note makes a narrower, testable claim, using a law drawn from the economics of evaluation under pressure. The shape of the reward hacking was set by the survival-like pressure of an impossible task with no rewarded way to fail — not by the grader. The fixes now floated are mostly grader fixes; a sharper grader relocates the behaviour without reducing it, unless it also lowers the pressure. The incident supplies an unusually clean case, including a grader that punished nothing yet drove a days-long attack against a check that did not exist. A 2×2 experiment crossing grader sharpness with a rewarded safe exit would falsify the claim if it is wrong.