arxivcs.CRcs.AIcs.LG2026-07-20
Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security
Devina Jain, David Hartmann, Chuan Li
LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluation, single-turn or multi-turn. We present a 21-scenario benchmark for \emph{adap…