AI Safety: Input Confidentiality vs. Output Correctness

Why “AI safety” means two different things — and why serious agents need both.