Supporting Artifact Evaluation with LLMs: A Study with Published Security Research Papers

This collection summarizes recent security and privacy research (multiple arXiv submissions) across reproducibility, identity, cryptography, ML-safety, and systems leakage. Key contributions include: an LLM-based artifact-evaluation toolkit that rates reproducibility (≈72% accuracy), autonomously prepares sandboxed execution for ~28% of runnable artifacts, and detects methodological pitfalls with F1>92%; a temporal jailbreak (TFM) for text-to-video models that converts unsafe prompts into sparse two-frame prompts to increase jailbreak success (up to +12%); discovery of system-level leakage for