Escaping the Linearity Trap: Manifold Detours for Black-Box Adversarial Attacks on Singing Audio Deepfake Detection

Collection of new research on adversarial and operational security for AI systems. Papers introduce novel black-box attacks (MARS against singing-voice deepfake detectors; DiffErase diffusion-based audio watermark removal), expose evaluation blindspots and surface-specific prompt injection vulnerabilities for tool-augmented LLM agents, and propose mitigations (dynamic per-request separators for Polymorphic Prompt Assembling). Other contributions include AdvScene for measuring real-world scene robustness of adversarial patches, CacheProbe auditing prompt-cache isolation in gateway APIs, an org‑