Known Unknowns: overcoming catastrophic failure modes in artificial intelligence