A Framework for Evaluating AI's Emerging Cyberattack Capabilities

Researchers propose a framework for evaluating AI's emerging cyberattack capabilities, identifying gaps in AI threat assessment through end-to-end attack chain analysis, helping defenders prioritize mitigations and conduct AI adversary simulation red teaming. The framework adapts an existing cyberattack chain framework to AI systems, analyzes real cases from more than 12,000 AI-involved cyber incidents collected by Google's Threat Intelligence Group, and distills seven representative attack chain archetypes. Through bottleneck analysis of these archetypes, the study locates the attack stages most susceptible to AI influence and uses externally developed cybersecurity model evaluations to focus on these key stages. The authors call this the most comprehensive AI cyber risk assessment framework published to date, report the potential for AI to amplify attack capabilities at specific attack stages, and provide defense priority recommendations.