Shared Vulnerabilities in Robustness-Optimized Defenses: One Breach Exposes the Family

This collection of recent arXiv papers highlights multiple active security risks and tooling advances across ML robustness, cryptanalysis, encrypted inference, LLM safety, and HPC agent security. Key findings: (1) Robustness-optimized defenses can share transferable vulnerabilities — a simple PGDTransfer attack yields ~80.4% average transfer success against purifier families at ε=4/255 — undermining purifier-based protection; (2) phishing classifiers that score >98% on clean data collapse under adversarial modifications (TF-IDF+LR and DistilBERT both fall to ~64%), showing real-world detection