Summary

When critical national infrastructure fails, or is coerced, during a crisis, it is of utmost importance to be able to quickly diagnose the source of the problem and fix it. While AI systems can support the recovery effort, they also create vulnerabilities, particularly where ownership of elements of that system is ambiguous.

This project builds a prototype toolkit for exploring AI opportunities and resilience to AI risks in crisis management planning. The toolkit consists of reusable, scenario-based tabletop exercises that let policymakers and stakeholders stress-test AI ownership dependencies and identify requirements for AI system integration.

Tyche is grounded on international institutional and academic literature on interdependence applied to the AI stack and to crisis decision-making, and draws on well-established scaffolding techniques for crisis management (e.g., NCSC Exercise in a box, CISA tabletop exercises).

Its novelty lies in treating AI dependency as a failure mode and to operationalise decision making for enabling AI opportunities and building resilience towards AI ownership risks in crisis management. Tyche’s focus on safe and secure AI for crisis management contributes to research in trustworthy AI.