E-MIA: Exam-Style Black-Box Membership Inference Attacks against RAG Systems

This collection of arXiv papers (May 5, 2026) presents multiple new offensive techniques and corresponding defenses that materially affect ML/AI security and software supply-chain risk. Key offensive contributions include E-MIA (an exam-style black-box membership inference attack that detects whether documents are present in RAG corpora), SRTJ (a self-evolving, training-free jailbreak framework that composes reusable attack rules), TC-UMIA (tri-class membership inference showing unlearning can leak retained data), and Trace (terminal-behavior fingerprinting of AI attack agents combined with a/