AI Security Policy Should Assess Systems, Not Only Models

Michael A. Riegler, Inga Strümke

We present swarm-attack, an open-source adversarial testing framework in which multiple lightweight LLM agents coordinate through shared memory, parallel exploration, and evolution…