AI Safety and Security Network (SSAIN)

Safety is traditionally concerned with harms in the physical world from cyber-physical systems (CPS) – loss of life, injury, or material/environmental damage. In contrast, cyber security is traditionally concerned with harmful digital effects from information technology (IT) systems – violation of privacy or unauthorised access to/corruption of sensitive information. Safety and security are often assessed independently, although their interdependence is recognised and there are some methods that analyse, for example, the safety impact of security violations, but this is surprisingly ill-explored. And AI adds further dimensions. First, the term “AI Safety” is now being used to include a broad range of harms (blurring the above distinctions), that also includes the unforeseen misuse of technology, unintended side-effects of increased autonomy and exploitation of the technology for malicious ends. Second, AI dramatically increases the attack surface (ways of compromising security) for digital systems. Third, the scope of concern of AI safety and security is much wider than traditional “safety-critical industries” or critical national infrastructure and, due to the pervasive nature of AI and its increased embedding in “everyday” applications, the issues need to be addressed in almost every domain that employs modern AI-based systems. Such issues cannot be resolved by a single research programme; instead, they are best addressed by cross-disciplinary collaborations bringing together understanding of real-world problems for developers, end users, policymakers, regulators, and leading research centres across arts & humanities, social science and computer science. The Safety and Security AI Network (SSAIN) will provide this collaborative environment, and support domain-specific implementation and validation of evolving principles, methods, e.g. algorithmic impact assessment (AIA) and tools for assessing AI safety and security. It will emphasise the fact that safety and security are not absolute and develop approaches that will support a holistic approach to safety and security risk management, reflecting the range of benefit-recipients and the risk-exposed individuals and groups. It will also consider the positive role of AI in safety and security, e.g. using AI to enhance the safety and security of citizens, and the use of AI in support of safety and security analysis. SSAIN will be set up as follows: An inclusive membership including academia, national institutes (e.g. the National Physical Laboratory (NPL) and the Alan Turing Institute) industry, policymakers, regulators, and standards bodies. A steering committee spanning relevant academic disciplines and representatives of the wider membership. Working groups on key issues, e.g. a research agenda for AI safety and security. Focused research and experimental work to support the working groups. Key outputs will include: A conceptual basis and prototypical analysis methods for assessing AI safety and security benefits and harms and for informing trade-offs between them. An inclusive and evolving community which can contribute to the responsible development and use of AI considering the potential safety and security benefits and harms of the technology. Exemplars of use of the analysis methods on representative use cases that can assist in their uptake, inform policy as well as help in refining the methods (likely in the automotive, healthcare, law enforcement and maritime domains). A research agenda for AI safety and security, reflecting both generic and domain-specific aspects. The network will focus on the UK, e.g. informing the AI standards hub, and supporting the government policy of responsible innovation, but also work with the wider international community where appropriate.