Misuse safeguards vs. loss-of-control safety

Two distinct categories of AI risk that are diverging in how well they are being addressed. Misuse safeguards prevent bad actors from using AI to do harm (e.g. generate malware or extremist content) — these are improving. Loss-of-control safety addresses the risk that an AI system pursues goals or takes actions its operators did not intend — this risk has been rising. They require completely different technical and governance approaches.

Boards that equate 'AI safety' solely with content filtering are missing the faster-growing risk: AI agents that act in ways humans did not sanction, cannot easily explain, or cannot shut down. The 51% rise in loss-of-control risk indicators over three quarters demands its own governance track.