15 AI Security Lessons From Black Hat and Ai4 2026

Executive summary: A Kiteworks survey and contemporaneous conference reporting show enterprises are rapidly deploying AI agents while governance and containment lag, and multiple high-impact technical threats were disclosed—most notably a GPU Rowhammer exploit against NVIDIA GDDR6 enabling host escalation, compiler optimization-induced race conditions, and a DPRK-linked malicious npm supply-chain insertion—alongside operational evidence that exploitation windows have collapsed to hours and that agent identity/privilege gaps are driving many incidents.