This seminar addresses the urgent and underexplored challenge of resilience in AI driven software systems – the ability of such systems to maintain or restore acceptable service levels under unexpected events, environmental shifts, and malicious threats. While societal awareness and regulatory frameworks (e.g., AI Act, GDPR, Cybersecurity Act) highlight the importance of resilience, the research landscape remains fragmented across machine learning (ML), software engineering (SE), and security. This Dagstuhl Seminar aims to bridge this gap by fostering a holistic, multidisciplinary dialogue on how resilience can be guaranteed by design.

The seminar’s novelty lies in its integrated perspective, uniting AI and SE experts to tackle resilience systematically. Three main questions will guide the discussions:

• What is resilience and how can it be assessed? We aim to move beyond narrow definitions (robustness, input sanitization) toward a comprehensive conceptual foundation that applies equally to algorithms and systems.
• How can AI algorithms be designed to be resilient? The seminar participants will identify fundamental limitations of current ML approaches, and explore new ways to evaluate and enhance model resilience.
• How can AI-driven software systems be engineered to be resilient by design? We will propose lifecycle touchpoints (e.g., in MLOps, SecDevOps) and blueprint development processes that embed resilience systematically.

Answering these questions promises multi-faceted impact:

• Scientific significance: No prior initiative has systematically combined ML foundations with SE methodologies to define, measure, and operationalize resilience.
• Practical relevance: Outputs such as benchmarks, case studies (e.g., AI code generation, self-driving perception), and guidelines will directly inform industry practices and future regulation.
• Societal impact: Failures of self-driving cars, chatbots, and large language models illustrate the real-world dangers of non-resilient AI. Ensuring resilience will mitigate risks of harm, misuse, and security breaches.

We target the following main outcomes:

• A unified definition of resilience across disciplines.
• Benchmarks and use cases to evaluate resilience comprehensively.
• A blueprint for resilient AI system development lifecycles.
• Foundations for joint publications, collaborations, and funding proposals.

Creative Commons BY 4.0

Axel-Cyrille Ngonga Ngomo, Eric Bodden, Christian Kästner, and Julia Rubin

This seminar qualifies for Dagstuhl's LZI Junior Researchers program. Schloss Dagstuhl wishes to enable the participation of junior scientists with a specialisation fitting for this Dagstuhl Seminar, even if they are not on the radar of the organizers. Applications by outstanding junior scientists are possible until Friday, December 12, 2025.