Catastrophic Risk Taxonomy

A warning is only as useful as the map that accompanies it. In “An Overview of Catastrophic AI Risks” (2023), Dan Hendrycks and collaborators did for AI danger what an epidemiologist does for disease: they refused to treat “risk” as a single undifferentiated cloud of dread and instead sorted it into distinct mechanisms, each with its own logic and its own potential remedies. The taxonomy organizes the threat into four families. The first is malicious use—bad actors turning AI toward harm, including engineered pandemics and cyberattacks. The second is the AI race—competitive pressure that systematically erodes safety because the actor who pauses falls behind. The third is organizational risk—the recognition that complex sociotechnical systems fail through accumulated human error and eroded safety margins. The fourth is rogue AI—systems that pursue misspecified objectives with increasing autonomy in ways nobody intended. The map’s value is not that it answers these questions but that it converts an undifferentiated terror into four tractable problems, each of which can be defended against by different means. You cannot