A central problem in AI safety is specifying objectives that truly reflect human preferences. Existing theory often quantifies how misspecification degrades performance, but it rarely separates mild sub‑optimality from truly catastrophic outcomes. To enable rigorous study of catastrophic risk from AI misalignment, this work introduces a formal definition of catastrophic performance. The definition relies on two reference values:\n\n- Baseline value: the utility achievable when no agent is deployed.\n- Orthogonal value: the utility of a trajectory that is unrelated to human preferences and, as described, is markedly worse than the baseline. The orthogonal value is typically far below the baseline.\n\nCatastrophic performance occurs when, after human activity has already improved the environment, an AI agent is introduced with the goal of further accelerating progress, yet its actions undo those human‑generated benefits, pushing outcomes toward the orthogonal value. Several historical examples illustrate the intuitive meaning of the baseline and orthogonal values, showing that catastrophic performance essentially represents a destructive rollback of human‑earned gains.\n\nReview: This definition offers a concrete benchmark for the worst‑case impact of AI misalignment, paving the way for quantitative risk metrics and safety mechanisms.