Systematic API Testing Through Model Checking and Executable Contracts

This collection highlights multiple software- and AI-security-relevant studies. Key findings: (1) PyTorch’s torch.compile produces silent correctness bugs that affect downstream LLM/DL behavior — the authors quantify this (19.2% of high-priority issues are incorrect outputs) and present AlignGuard, an LLM-driven mutation fuzzer that found 23 new correctness bugs (14 marked high-priority). (2) DeepGuard proposes a multi-layer representation aggregation and training/inference steering to reduce insecure code generation from code LLMs, improving secure-and-correct generation rates over strong bas