Programme Risk Day 2026

Abstracts

Karina Schreiber
Building Risk Culture to Independently Challenge Opinions – Including the Machine's

Risk culture is often called “soft” - until it’s the difference between a near‑miss and a headline. This talk makes risk culture tangible: the behaviours that drive better decisions - cognitive diversity, constructive dissent, escalation, and integrity in execution. Then we bring those foundations into the AI era, where speed and opacity can amplify both good judgment and small errors.
 

Michael Bisig
From Zero Trust to AI Agents: How Digital Identity and Trust are becoming the Core of IT Security (and Business)

In today’s fast-evolving and interconnected world, the pace of change is accelerating. Flexible work models are now the norm, processes are increasingly digitized and automated, and AI-driven agents are revolutionizing workflows. These forces are reshaping the foundations of both business operations and IT systems. This talk explores the central role of digital identity and trust in shaping this transformation. As organizations embrace modern technologies, digital identities have emerged as the new security perimeter and have become the control plane for business, offering new opportunities and posing new risks.
 

Olga Fink
Scalable and Trustworthy Industrial AI: From Graph Learning and Physics-Informed AI to Foundation Models and Agentic AI

Industrial AI is undergoing a fundamental transition from task-specific machine learning models toward general-purpose AI systems capable of learning, reasoning, and supporting autonomous decision-making across diverse engineering applications. However, industrial systems pose unique challenges arising from sparse labels, complex spatiotemporal dependencies, heterogeneous sensor networks, evolving operating conditions, and strict requirements on reliability, interpretability, and safety. Addressing these challenges requires AI systems that move beyond purely data-driven prediction by integrating physical knowledge, engineering structure, and autonomous reasoning.

This talk presents recent advances toward scalable Industrial AI along four complementary directions. First, spatiotemporal graph neural networks (GNNs)exploit the topology and interactions of engineering systems to model spatial and temporal dependencies, enabling improved monitoring, forecasting, virtual sensing, and fault detection across interconnected assets. Second, physics-informed learning embeds physical inductive biases, such as governing equations, conservation laws, symmetries, and domain knowledge, into machine learning models, enabling more data-efficient, physically consistent, and robust learning with improved extrapolation beyond observed operating conditions. Third, tabular foundation models introduce a new paradigm based on in-context learning, enabling a single pretrained model to solve diverse industrial prediction tasks with little or no task-specific retraining. This opens new opportunities for scalable industrial AI, while also raising important questions about how structured engineering data, including temporal signals and system-level dependencies, should be represented for foundation-model-based inference. Finally, agentic AI extends predictive models toward intelligent engineering assistants by combining foundation models with graph representations, digital twins, external tools, and domain knowledge to perform autonomous root-cause analysis, evidence-based reasoning, and interactive diagnosis of industrial systems.

Drawing on applications in intelligent maintenance, energy systems, and industrial process monitoring, the talk illustrates how these complementary paradigms collectively enable the next generation of scalable and trustworthy Industrial AI. By integrating graph-based representations, physics-informed learning, foundation models, and agentic reasoning, future AI systems will combine predictive intelligence with physical understanding and reasoning capabilities, paving the way toward intelligent engineering assistants that support diagnosis, root-cause analysis, and, ultimately, autonomous decision-making in complex engineering systems.
 

Arthur Gervais
Would you hire an AI as your next Security Engineer?

Cybersecurity teams face a growing mismatch between the scale of their responsibilities and the availability of experienced security engineers. Could an organization address this gap by hiring an AI employee?

This talk examines whether an AI employee is ready to fill a security engineering role while potentially working with CISOs and existing human teams. Unlike conventional scanners or assistants, an AI security engineer can demonstrate agency: investigating applications, formulating and testing hypotheses, adapting its approach based on evidence, identifying vulnerabilities, and producing reports, code, and configurations.

Drawing on lessons from building BugBunny, I will consider which responsibilities an AI can already assume, where human expertise and judgment remain essential, and how work should be divided between the AI, security engineers, and the CISO. The discussion will examine authorization, reliability, oversight, evidence, accountability, and trust. Rather than assuming that an AI can - or should - replace a human engineer, the talk asks what capabilities and safeguards would need to be in place before an organization could confidently make such a hire.
 

Rama Cont
tba
 

Daniel Grosshans
Extreme Risk in High-Dimensional Portfolios 

Measuring risk in a high-dimensional portfolio is hard: available methods are either overly simplistic - unable to capture tail dependence in severe downturns - or computationally prohibitive due to the curse of dimensionality. We present an architecture that computes portfolio risk fast and with guaranteed accuracy. It separates the joint law of portfolio constituents into fast-changing marginals, updated online as markets move, and a slow-moving dependence layer, compiled offline from history. The approach is simulation-free, scales polynomially in the portfolio dimension, and carries uniform error bounds with a controllable runtime–accuracy trade-off. The guarantees hold deep in the tail, where accuracy matters most and simulation is least reliable, and the output is a full joint distribution consistent with its inputs by construction. For large portfolios with specific downside risks, like a pension fund keeping its funding ratio above a critical threshold, this opens new avenues for targeted risk management.