Scaling AI Safety Service for a Multi-Agent World

The agency is seeking a scalable AI safety service designed for a multi-agent world. The service must be scalable to a realistic number of agents and diverse agentic tools and knowledge bases. It needs to capture high-fidelity behaviors of frontier AI agents and be externally valid, allowing for principled characterization of simulation-derived conclusions. The service must also be safe and secure to study dangerous collective behaviors without risk of uncontrolled deployment, and reproducible to enable comparison of different methods. The project covers developing principled methods to use real-world deployment data for designing grounded environments and evaluating their external validity. It also includes evaluating vulnerabilities in networks of AI agents and developing new evaluation frameworks for risks specific to multi-agent deployments, such as resilience to adversarial sub-populations, propagation of attacks between agents, and susceptibility to cascading failures. Eligibility requires bidders to be a USA Organization Only. Performance of the work will be offsite. The agency is looking for proposals.