Google DeepMind has developed a comprehensive framework to evaluate cybersecurity risks posed by advanced AI systems. The initiative addresses growing concerns about AI's dual-use potential—enhancing defenses while enabling sophisticated attacks. By analyzing 12,000 real-world cyberattacks across 20 countries, the team identified seven attack archetypes, including phishing and malware deployment, and pinpointed critical vulnerabilities in the cyberattack chain where AI could disrupt traditional defense costs.

The framework adapts established cybersecurity evaluation models like MITRE ATT&CK to assess AI's ability to automate attacks across 50 technical challenges spanning reconnaissance to system infiltration. Early tests suggest current AI models lack breakthrough offensive capabilities in isolation, but the team warns that as AI advances, attack methods will evolve. Notably, the evaluation highlights overlooked attack phases like evasion and persistence, where AI could lower operational barriers for threat actors.

Cybersecurity teams can use the benchmark to simulate AI-powered attacks during red teaming exercises and prioritize defenses. The framework emphasizes proactive risk mitigation over reactive measures, urging organizations to update protections as AI systems scale. Google's research underscores the urgency of balancing AI's cybersecurity benefits with its potential misuse, particularly as artificial general intelligence approaches practical implementation.

This technical roadmap provides actionable insights for defenders while establishing a baseline for evaluating future AI systems. By focusing on real-world attack patterns and system vulnerabilities, the framework aims to foster responsible AGI development alongside robust cybersecurity protocols.