Attackers exploit flaws in about five days; autonomous AI agents exploited 87% of one-day flaws unaided in peer-reviewed tests.