Frontier AI systems, the most advanced AI models, now act autonomously over long tasks with little oversight. They may go beyond effective human control, and by the time that is obvious it can be too late to prevent irreversible harm. Such risk must be predicted before deployment.
The framework treats loss of control as the joint effect of misaligned motive (goals diverging from human intent), harm-enabling capability (ability to cause harm), and monitoring evasion (ability to keep acting despite oversight). These are measured with evaluation tasks and combined into a single loss-of-control score.
Across 13 frontier models, that score strongly predicted failure rates in contained long-horizon scenarios (correlation 0.84). Risk was not just capability. Mitigations targeted at diagnosed weaknesses reduced failures while preserving general performance, making loss of control measurable and manageable.