HySafe-AI: Hybrid Safety Architectural Analysis Framework for AI Systems: A Case Study

Traditional safety methods need updating for modern AI: The paper demonstrates that conventional safety analysis techniques like FMEA (Failure Modes and Effects Analysis) and FTA (Fault Tree Analysis)—which have long protected critical systems in aviation and manufacturing—are insufficient for end-to-end AI architectures such as large language models and vision systems, requiring fundamental adaptations to address how these systems work internally.

"Black box" AI poses unique safety challenges: The research highlights that foundational AI models create safety analysis problems through their use of "latent representations" (internal patterns the AI learns that aren't directly interpretable), making it difficult to predict failure modes using traditional methods designed for systems with transparent, rule-based logic.

Hybrid approaches offer a path forward: HySAFE-AI combines traditional safety engineering practices with new methods tailored to AI's unique characteristics, suggesting that the solution isn't abandoning proven safety techniques but rather adapting them—a practical middle ground for evaluating autonomous vehicles, robots, and other safety-critical AI applications.

Safety standards for AI are urgently needed: The paper emphasizes that as AI systems increasingly control critical infrastructure and autonomous vehicles, the field lacks standardized safety evaluation frameworks, and current industry practices are inadequate for protecting public safety.