Signal Reducing Catastrophic Risk from AI with Systematic Monitoring and Evaluation of Rogue AI Progression

Summary

This paper presents a structured framework of behavioral indicators intended to signal when an AI system may be progressing toward a potentially catastrophic threat. Its approach is pragmatic, drawing inspiration from methodologies already established in cybersecurity and national security. The framework sets clear metrics, indicators, and thresholds across multiple dimensions of AI capability and behavior. The goal is to let researchers and policymakers implement monitoring protocols that are grounded in concrete evidence rather than intuition. Its central feature is adapting threat-signal thinking from cybersecurity into the domain of AI risk monitoring.