Experimenting with AI to defend critical infrastructure (red.anthropic.com)

In a groundbreaking collaboration, Anthropic and the Pacific Northwest National Laboratory (PNNL) are leveraging AI to enhance the defense of critical infrastructure against cyber attacks. By utilizing Claude, an advanced AI model, researchers successfully simulated cyber attacks on a water treatment plant in just three hours, significantly reducing the time typically required by human experts for such red teaming exercises. This proof of concept highlights the potential of AI to rapidly identify vulnerabilities and improve system defenses, demonstrating a vital step forward in public-private partnerships aimed at bolstering national security. The project focused on adversary emulation, allowing the AI to model specific threat actors and attacks to better understand system weaknesses. PNNL developed a "scaffold" that enabled Claude to translate natural language prompts into complex attack chains, which facilitated the identification of blind spots within the cyber-physical systems. Impressively, Claude exhibited resourcefulness by switching to different attack methods when initial attempts failed, showcasing how improving AI capabilities can aid both defenders and attackers. This research is a crucial development in the dual-use nature of AI, emphasizing the importance of innovative defenses amid increasing cyber threats.