AI for Science, Software and Security

The complex relationship between the three S, science, software, and security (S3) was first discussed. It was categorized into four distinct domains: "Software of Science", where software tools are used to facilitate scientific endeavors, such as data analysis and computational modeling tools; "Security of Science", which focuses on the protection of scientific data and systems from unauthorized access and tampering, for example the novel field of cyberbiosecurity to safeguard biological data; "Science of Software", where scientific methodologies are applied to enhance software quality; and "Science of Security", such as cryptography, which leverages scientific principles to tackle complex security challenges.

The workshop identified a set of challenges posed by the intersection of AI and software development, specifically in seamlessly integrating AI-generated software with traditional frameworks: knowledge transfer between domains, balancing accuracy and scalability, ensuring trustworthiness as well as reliability and safety. To overcome these challenges, it was proposed that efforts should focus on supporting domain experts in computational science, enhancing explainability and accountability in AI systems, and exploring low-code or no-code solutions and assisted code development. 

In terms of the intersection of AI and security, the complexities of tailoring AI systems for specific use cases were discussed. Challenges were identified, including AI limitations compared to human intelligence, potential errors, design-level challenges, and domain specialization through fine-tuning AI whilst respecting privacy, safety, and legal considerations during data acquisition. The discussion underscored the absence of standardized metrics akin to straightforward scientific measurements, advocating for the establishment of validation mechanisms for AI correctness and the integration of human-AI collaboration with the ability to deactivate systems when necessary. 

The workshop highlighted the potential of AI to enhance scientific endeavors by facilitating hypothesis testing, improving question quality, and boosting research productivity, while emphasizing the importance of ethical guidelines and robust validation processes to mitigate misuse and bias. Participants also emphasized democratizing access to formal analysis tools across scientific domains using AI, despite challenges such as data scarcity, resource limitations, and data quality issues. They proposed a strategic focus on narrower domains like security protocols and outlined the need for clear roadmaps to navigate hurdles in data acquisition and utilization, ensuring scalability, safety, and correctness in formal methods through increased computational resources and skilled personnel.