AI to defend critical infrastructure

**Executive summary:** Anthropic partnered with PNNL to use the Claude model to accelerate adversary emulation against a high-fidelity water treatment plant simulation, demonstrating faster attack reconstruction (hours vs. weeks), automated attack-chain generation, and model resourcefulness such as switching UAC bypass techniques when a predefined tool failed; the work is presented as defensive research and a model for public–private collaboration to improve critical infrastructure security.