Cascade: Composing Software-Hardware Attack Gadgets for Adversarial Threat Amplification in Compound AI Systems

arXiv 2603.12023·high signal

Introduces a CVE-style taxonomy of attack gadgets composable across the traditional software and hardware stack underlying Compound AI pipelines (LLMs + tools + databases). Demonstrates that classic CVE-documented software flaws, combined with hardware-level side-channels, create amplified threat surfaces unique to multi-model AI systems. Directly actionable: builders of LLM pipelines must treat every dependency layer—not just model inputs—as an attack surface.