arXiv cs.CLPaper
Layered LLM Defenses as an Ensemble: Access Tiers, Inference Cost, and the Measured Failure Correlation Between Defense Layers
This is the kind of unglamorous but critical work that prevents you from shipping false confidence in your safety stack. Most teams assume that adding a second defense halves attack success; this paper shows it often doesn't. If you're building security into an LLM product, read this and rethink your defense architecture.