The 89th meeting of IFIP Working Group 10.4 was held at Vytautas Magnus University in Kaunas, Lithuania, under the theme “The Past Shapes the Future.” The event brought together pioneers of dependable computing alongside researchers tackling today’s emerging challenges.
The opening session honoured Al Avižienis, Jean-Claude Laprie, Hermann Kopetz, Brian Randell, John Meyer and Bill Carter for their foundational contributions to dependable computing. Building on their legacy, the technical sessions examined the current state of the field and explored the challenges likely to shape its future.
While the discussions spanned multiple aspects of dependable computing, five themes emerged repeatedly as priorities for future research and collaboration.
Artificial Intelligence becomes the defining challenge
Artificial Intelligence (AI) is no longer a peripheral topic for dependable computing. It is rapidly becoming the defining challenge for the discipline. Participants agreed that agentic AI systems are already being deployed in critical domains, and that the risks of disengagement now outweigh the risks of engagement.
AI systems, particularly large language models (LLMs) and agentic architectures, introduce failure modes that fall outside traditional dependability frameworks. Their stochastic behaviour, sensitivity to distributional shifts, and the often opaque relationship between intended and actual behaviour make assurance significantly more difficult. Iterative agentic loops can degrade performance, sequential tool chains can amplify otherwise benign actions, and evaluations based solely on accuracy are inadequate. Dependability assessments must also consider consistency, robustness, predictability and safety.
Rather than abandoning established approaches, participants argued that the core principles of dependable computing should be adapted for AI systems. Fault tolerance, redundancy, formal specification and structured assurance remain essential. In safety-critical domains such as aviation, automotive systems and critical infrastructure, AI should be introduced with the same rigour applied to previous disruptive technologies through independent assessment, staged validation and assurance levels proportional to the consequences of failure.
Understandability should become a first-class property
Systems are increasingly being deployed without anyone, including developers, operators or regulators, fully understanding how they behave. Although this is not a new problem, AI-generated code is accelerating the trend by enabling the creation of software that may not have been fully reasoned about by any individual.
Participants argued that understandability should become a first-class research and engineering objective rather than a secondary concern, particularly when code is reused. Systems that clearly express their designers’ intent are easier to reason about, verify and audit when failures occur.
This perspective favours architectures that preserve decomposability and auditability over monolithic end-to-end designs, even where the latter may deliver short-term performance gains. Likewise, assurance cases should evolve from static documents into living frameworks that remain valid as systems continue to change.
Hardware can no longer be assumed dependable
Hardware can no longer be assumed to provide a dependable foundation for software. At modern transistor scales, crosstalk, environmental noise and side-channel vulnerabilities are inherent characteristics rather than defects that can simply be engineered away. Increasingly complex predictive execution mechanisms, together with opaque supply chains for trusted execution environments, further weaken assumptions about secure hardware boundaries.
As a result, dependability must be reconsidered from the ground up, treating hardware as a distributed system in its own right. Chiplet-based architectures, which organise processors as networks of small, independently programmable functional units, were identified as a promising approach for improving fault tolerance and physical isolation. An important open question is the level at which fault and intrusion tolerance should be implemented, whether at the node, chip, chiplet or functional unit level.
Dependable computing must engage with policy and legislation
Much of the technical knowledge needed to build more dependable systems already exists. What is missing is a set of regulatory and market incentives that encourage its application.
Two well-documented cases illustrate that even catastrophic failures have not produced effective legislative responses. One involved forensic software whose false-positive error rate was presented in court as one in a million but was later shown through rigorous analysis to be closer to one in eight. The second was the British Post Office Horizon scandal, in which hundreds of innocent people were wrongly prosecuted on the basis of faulty software.
Participants argued that this represents a structural failure that the dependability community must address directly. Software vendors face little accountability, programmers are generally uncertified, and courts often lack the technical expertise needed to evaluate software evidence rigorously. Greater engagement with policymakers, stronger standards for software testing and validation, and regulatory frameworks that create meaningful incentives for dependability are all needed. Experience from building codes, automotive safety and aviation certification suggests that lasting change requires both incentives and penalties, often triggered by major failures.
Intelligent vehicles define a near-term research agenda
Self-driving vehicles are entering large-scale deployment, but participants agreed they have not yet demonstrated reliable autonomous operation, even within constrained operational domains. The gap between claimed and actual autonomy remains significant.
Several opportunities exist for IFIP WG 10.4 to make important contributions during the 2026 to 2030 period, including defining quantitative acceptance criteria for AI and machine learning systems, improving the collection of field data on incidents and tele-assistance interventions, developing comprehensive system assurance cases, and recommending technologies that simplify assurance activities.
Looking beyond 2030, attention is expected to shift from individual autonomous vehicles to Systems of Autonomous Systems that integrate self-driving cars, autonomous transport and service vehicles, unmanned aerial vehicles, and intelligent infrastructure. Ensuring the dependability of these interconnected ecosystems will require new approaches to acceptance criteria, assurance case development and system architecture, creating a substantial long-term research agenda for the dependable computing community.